A build fix wrt gevent-1.5 + benchmarks for nogil go and channels.

Starting from gevent >= 1.5 '*.pxd' files for gevent API are no longer
provided, at least in released gevent wheels. This broke pygolang:

    Error compiling Cython file:
    ------------------------------------------------------------
    ...

    # Gevent runtime uses gevent's greenlets and semaphores.
    # When sema.acquire() blocks, gevent switches us from current to another greenlet.

    IF not PYPY:
        from gevent._greenlet cimport Greenlet
       ^
    ------------------------------------------------------------

    golang/runtime/_runtime_gevent.pyx:28:4: 'gevent/_greenlet.pxd' not found

Since gevent upstream refuses to restore Cython level access[1], let's fix the
build by using gevent bits via Python-level.

Even when used via py import gevent-1.5 brings speed improvement compared to
gevent-1.4 (used via cimport):

	(on i7@2.6GHz, gevent runtime)

                      gevent-1.4   gevent-1.5
                      (cimport)    (py import)

    name              old time/op  new time/op  delta
    pyx_select_nogil  9.47µs ± 0%  8.74µs ± 0%   -7.70%  (p=0.000 n=10+9)
    pyx_go_nogil      14.3µs ± 1%  12.0µs ± 1%  -16.52%  (p=0.000 n=10+10)
    pyx_chan_nogil    7.10µs ± 1%  6.32µs ± 1%  -10.89%  (p=0.000 n=10+10)
    go                16.0µs ± 2%  13.4µs ± 1%  -16.37%  (p=0.000 n=10+10)
    chan              7.50µs ± 0%  6.79µs ± 0%   -9.53%  (p=0.000 n=10+10)
    select            10.8µs ± 1%  10.0µs ± 1%   -6.78%  (p=0.000 n=10+10)

Using gevent-1.5 could have been even faster via cimport (it is still
possible to compile and test against gevent installed in development
mode via `pip install -e` because pxd files are there in gevent worktree
and tarball):

                      gevent-1.5   gevent-1.5
                      (py import)  (cimport)

    name              old time/op  new time/op  delta
    pyx_select_nogil  8.74µs ± 0%  7.90µs ± 1%  -9.60%  (p=0.000 n=9+10)
    pyx_go_nogil      12.0µs ± 1%  11.2µs ± 2%  -6.35%  (p=0.000 n=10+10)
    pyx_chan_nogil    6.32µs ± 1%  5.89µs ± 0%  -6.80%  (p=0.000 n=10+9)
    go                13.4µs ± 1%  12.4µs ± 1%  -7.54%  (p=0.000 n=10+9)
    chan              6.79µs ± 0%  6.42µs ± 0%  -5.47%  (p=0.000 n=10+10)
    select            10.0µs ± 1%   9.4µs ± 1%  -6.39%  (p=0.000 n=10+10)

but we cannot use cimport to access gevent-1.5 universally, since pxd are not
shipped in gevent wheel releases.

In the future we might want to change plain version check into compile time
check whether gevent/_greenlet.pxd is actually present or not and use faster
access if yes. Requesting gevent to be installed in non-binary form
might be also an option worth trying.

However plain version check should be ok for now.

[1] https://github.com/gevent/gevent/issues/1568

on i7@2.6GHz it looks like:

thread runtime:

    name              time/op
    pyx_select_nogil  2.70µs ±13%
    pyx_go_nogil      15.9µs ± 1%
    pyx_chan_nogil    2.79µs ± 2%
    go                17.6µs ± 0%
    chan              3.05µs ± 4%
    select            3.62µs ± 4%

gevent runtime (gevent-1.4.0):

    name              time/op
    pyx_select_nogil  9.39µs ± 1%
    pyx_go_nogil      15.1µs ± 2%
    pyx_chan_nogil    7.10µs ± 1%
    go                16.6µs ± 1%
    chan              7.47µs ± 1%
    select            10.7µs ± 0%

Just a single change to expose time constants to C++ users (2476f47e).

This makes them available for C++ users as well.

On macos and windows, Python2 is built with --enable-unicode=ucs2, which
makes it to use UTF-16 encoding for unicode characters, and so for
characters higher than U+10000 it uses surrogate encoding with _2_
unicode points, for example:

        >>> import sys
        >>> sys.maxunicode
        65535                       <-- NOTE indicates UCS2 build
        >>> s = u'\U00012345'
        >>> s
        u'\U00012345'
        >>> s.encode('utf-8')
        '\xf0\x92\x8d\x85'
        >>> len(s)
        2                           <-- NOTE _not_ 1
        >>> s[0]
        u'\ud808'
        >>> s[1]
        u'\udf45'

This leads to e.g. b tests failing for

    # tbytes                        tunicode
    (b"\xf0\x90\x8c\xbc",           u'\U0001033c'),     # Valid 4 Octet Sequence '𐌼'

    >           assert b(tunicode) == tbytes
    E           AssertionError: assert '\xed\xa0\x80\xed\xbc\xbc' == '\xf0\x90\x8c\xbc'
    E             - \xed\xa0\x80\xed\xbc\xbc
    E             + \xf0\x90\x8c\xbc

because on UCS2 python build u'\U0001033c' is represented as 2 unicode
points:

    >>> s = u'\U0001033c'
    >>> len(s)
    2
    >>> s[0]
    u'\ud800'
    >>> s[1]
    u'\udf3c'
    >>> s[0].encode('utf-8')
    '\xed\xa0\x80'
    >>> s[1].encode('utf-8')
    '\xed\xbc\xbc'

-> Fix it by detecting UCS2 build and working around by manually
combining such surrogate unicode pairs appropriately.

A reference on the subject:

https://matthew-brett.github.io/pydagogue/python_unicode.html#utf-16-ucs2-builds-of-python-and-32-bit-unicode-code-points

This is a preparatory step for the next patch where we'll be fixing
strconv for Python2 builds with --enable-unicode=ucs2, where a unicode
character can be taking _2_ unicode points.

In that general case relying on unicode objects to represent runes is
not good, because many things generally do not work for U+10000 and
above, e.g. ord breaks:

    >>> import sys
    >>> sys.maxunicode
    65535                       <-- NOTE indicates UCS2 build
    >>> s = u'\U00012345'
    >>> s
    u'\U00012345'
    >>> s.encode('utf-8')
    '\xf0\x92\x8d\x85'
    >>> len(s)
    2                           <-- NOTE _not_ 1
    >>> ord(s)
    Traceback (most recent call last):
      File "<stdin>", line 1, in <module>
    TypeError: ord() expected a character, but string of length 2 found

so we switch to represent runes as integer, similarly to what Go does.

Commit 8af78fc5 (pyx.build: v↑ setuptools_dso (1.2 -> 1.4)) upgraded
setuptools_dso to 1.4, but since from

https://github.com/mdavidsaver/setuptools_dso/commit/3f3ff746

setuptools_dso started to use multiprocessing, pyx.build, when running
under gpython, started to hang, which is a known gevent problem - see
e.g. here: https://github.com/gevent/gevent/issues/993. The problem was
manifesting itself as pyx.build unit test hanging under Python3.

Fix it by installing gevent multiprocessing plugin which is
automatically used/activated by gevent.monkey.patch_all().

geventmp says it is pre-alpha, but by using it we can unhang pyx.build
tests, which is better state than before. The other future possibility
would be to use https://github.com/jgehrcke/gipc wrapped into
multiprocessing compatible API.

This is top-level documentation for error chaining that was promised
and marked as TODO in

- fd95c88a (golang, errors, fmt: Error chaining (C++/Pyx))
- 17798442 (golang: Expose error at Py level)
- 78d0c76f (golang: Teach pyerror to be a base class)
- 337de0d7 (golang, errors, fmt: Error chaining (Python))
- 03f88c0b (errors: Take .__cause__ into account)

This provides top-level documentation for b and u that was promised and
marked as TODO in bcb95cd5 (golang: Provide b, u for strings).

Pychan provides __eq__ (see 2c8063f4 "*: Channels must be compared by
==, not by "is" even for nilchan"), but does not provide __ne__. At the
same time in 17798442 (golang: Expose error at Py level) we had to
define both pyerror.__eq__ and pyerror.__ne__ because without __ne__
pyerror != pyerror was not working correctly.

As it turns out pychan != pychan already works ok, because pychan does
not have base class and for that case cython automatically generates
__ne__ based on __eq__:

    https://github.com/cython/cython/blob/0.29.14-629-ga73815042/Cython/Compiler/ModuleNode.py#L1963-L1976
    https://github.com/cython/cython/commit/b75d2942afab

Add corresponding comment and extend tests to make sure it is indeed so.

Go version does not provide this, but the topic of sync.RWMutex
downgrading was raised up several times, at least

	https://github.com/golang/go/issues/4026
	https://github.com/golang/go/issues/23513
	https://groups.google.com/forum/#!topic/golang-nuts/MmIDUzl8HA0
	...

Atomic downgrading is often useful to avoid race window in between
Unlock and RLock and, as consequence, having the need to recheck things
after RLock.

We can put this complexity and logic into well-defined RWMutex primitive
instead of throwing it to be solved by every RWMutex user.

Provide sync.RWMutex that can be useful for cases when there are
multiple simultaneous readers and more seldom writer(s).

This implements readers-writer mutex with preference for writers
similarly to Go version.

Only io.EOF and io.ErrUnexpectedEOF for now.
Moved here from wcfs from wendelin.core.

A Python error can have links to other errors by means of both .Unwrap()
and .__cause__ . These ways are both explicit and so should be treated
by e.g. errors.Is as present in error's error chain.

It is a bit unclear, at least initially, how to linearise and order
error chain traversal in divergence points - for exception objects where
both .Unwrap() and .__cause__ are !None. However more closer look
suggests linearisation rule to traverse into .__cause__ after going
through .Unwrap() part - please see details in documentation added into
_error.pyx

-> Teach errors.Is to do this traversal, and this way now e.g. exception
raised as

	raise X from Y

will be treated by errors.Is as being both X and Y, even if any of X or Y
also has its own error chain via .Unwrap().

Top-level documentation is TODO.

Following errors model in Go and fd95c88a (golang, errors, fmt: Error
chaining (C++/Pyx)) let's add support at Python-level for errors to wrap
each other and to be inspected/unwrapped:

- an error can additionally provide way to unwrap itself, if it
  provides .Unwrap() method. .__cause__ is not taken into account yet,
  but will be in a follow-up patch;
- errors.Is(err) tests whether an item in error's chain matches target;
- `fmt.Errorf("... : %w", ... err)` is similar to `"... : %s" % (..., err)`
  but resulting error, when unwrapped, will return err.
- errors.Unwrap is not exposed as chaining through both .Unwrap() and
  .__cause__ will need more than just "current element" as unwrapping
  state (i.e. errors.Unwrap API is insufficient - see next patch), and
  in practice users of errors.Unwrap() are very seldom.

Support for error chaining through .__cause__ will follow in the next
patch.

Top-level documentation is TODO.

See https://blog.golang.org/go1.13-errors for error chaining overview.

It is surprising to have an exception class that cannot be derived from.

Besides, in the future we'll use subclassing from golang.error as an
indicator that an error is a "well-defined" (in simple words - does not
need traceback to be interpreted).

The first step to expose errors and error chaining to Python:

- Add pyerror that wraps a pyx/nogil C-level error and is exposed as golang.error at py level.
- py errors must be compared by ==, not by "is"
- Add (py) errors.New to create a new error from text.
- a C-level error that has .Unwrap, is exposed with .Unwrap at py level,
  but full py-level chaining will be implemented in a follow-up patch.
- py error does not support inheritance yet.

Top-level documentation is TODO.

Following errors model in Go, let's add support for errors to wrap other
errors and to be inspected/unwrapped:

- an error can additionally provide way to unwrap itself, if it
  implements errorWrapper interface;
- errors.Unwrap(err) tries to extract wrapped error;
- errors.Is(err) tests whether an item in error's chain matches target;
- `fmt.errorf("... : %w", ... err)` is similar to `fmt.errorf("... : %s", ... err.c_str())`
  but resulting error, when unwrapped, will return err.

Add C++ implementation for the above + tests.
Python analogs will follow in the next patches.

Top-level documentation is TODO.

See https://blog.golang.org/go1.13-errors for error chaining overview.

Package cxx was added in 9785f2d3 (cxx: New package), but the interface
that cxx:dict provided turned out to be not optimal:

    dict.get  was returning (v, ok), and
    dict.pop  ----//---

Correct dict.get and dict.pop to return just value, and, similarly to
channels API, provide additional dict.get_ and dict.pop_ - extended
versions that also return ok:

    dict.get(k)  -> v
    dict.pop(k)  -> v
    dict.get_(k) -> (v, ok)
    dict.pop_(k) -> (v, ok)

This time add tests.

Follow the scheme established and used for all other packages, because
we will soon have fmt pyx part which, if named as fmt.pyx, will
intersect and conflict with fmt.py .

errors.New was added in a245ab56 (errors: New package) without test.

Makes understanding which test is it and where when one fails.

Currently libgolang_test.cpp contains tests for code in libgolang.cpp and
for code that lives in other libgolang packages - sync, fmt, etc. It is
becoming tight and we are going to split libgolang_test.cpp and move
package tests to their corresponing files - e.g. to sync_test.cpp and
the like.

Move common assertion utilities into shared header before that as a
preparatory step.

Just use builtins and cimported things that we have at pyx level.

U is preffered way to make sure an object is unicode string.

This will allow to integrate qq with u in the next patch.

Moving to compiled code for string processing functions is also
generally better for performance.

With Python3 I've got tired to constantly use .encode() and .decode();
getting exception if original argument was unicode on e.g. b.decode();
getting exception on raw bytes that are invalid UTF-8, not being able to
use bytes literal with non-ASCII characters, etc.

So instead of this pain provide two functions that make sure an object
is either bytes or unicode:

- b converts str/unicode/bytes s to UTF-8 encoded bytestring.

	Bytes input is preserved as-is:

	   b(bytes_input) == bytes_input

	Unicode input is UTF-8 encoded. The encoding always succeeds.
	b is reverse operation to u - the following invariant is always true:

	   b(u(bytes_input)) == bytes_input

- u converts str/unicode/bytes s to unicode string.

	Unicode input is preserved as-is:

	   u(unicode_input) == unicode_input

	Bytes input is UTF-8 decoded. The decoding always succeeds and input
	information is not lost: non-valid UTF-8 bytes are decoded into
	surrogate codes ranging from U+DC80 to U+DCFF.
	u is reverse operation to b - the following invariant is always true:

	   u(b(unicode_input)) == unicode_input

NOTE: encoding _and_ decoding *never* fail nor loose information. This
is achieved by using 'surrogateescape' error handler on Python3, and
providing manual fallback that behaves the same way on Python2.

The naming is chosen with the idea so that b(something) resembles
b"something", and u(something) resembles u"something".

This, even being only a part of strings solution discussed in [1],
should help handle byte- and unicode- strings in more robust and
distraction free way.

Top-level documentation is TODO.

[1] nexedi/zodbtools!13

This continues 60f6db6f (libgolang: Provide nil as alias for nullptr and
NULL): I've tried to compile pygolang with Clang on my Debian 10
workstation and got:

    $ CC=clang CXX=clang++ python setup.py build_dso -i

    In file included from ./golang/fmt.h:32:
    ./golang/libgolang.h:381:11: error: unknown type name 'nullptr_t'; did you mean 'std::nullptr_t'?
    constexpr nullptr_t nil = nullptr;
              ^~~~~~~~~
              std::nullptr_t
    /usr/bin/../lib/gcc/x86_64-linux-gnu/8/../../../../include/x86_64-linux-gnu/c++/8/bits/c++config.h:242:29: note: 'std::nullptr_t' declared here
      typedef decltype(nullptr)     nullptr_t;
                                    ^
    :
    In file included from ./golang/context.h
    In file included from golang/runtime/libgolang.cpp:30:
    ./golang/libgolang.h:381:11: error: unknown type name 'nullptr_t'; did you mean 'std::nullptr_t'?
    constexpr nullptr_t nil = nullptr;
              ^~~~~~~~~
              std::nullptr_t
    /usr/bin/../lib/gcc/x86_64-linux-gnu/8/../../../../include/x86_64-linux-gnu/c++/8/bits/c++config.h:242:29: note: 'std::nullptr_t' declared here
      typedef decltype(nullptr)     nullptr_t;
                                    ^
    :39:
    ./golang/libgolang.h:381:11: error: unknown type In file included from golang/fmt.cpp:25:
    In file included from ./golang/fmt.h:32:
    ./golang/libgolang.h:421:17: error: unknown type name 'nullptr_t'; did you mean 'std::nullptr_t'?
        inline chan(nullptr_t) { _ch = nil; }
                    ^~~~~~~~~
                    std::nullptr_t

    ...

It seems with GCC and Clang under macOS nullptr_t is automatically provided in
builtin namespace, while with older Clang on Linux (clang version 7.0.1-8) only
in std:: namespace - rightfully as nullptr_t is described to be present there:

https://en.cppreference.com/w/cpp/types/nullptr_t

This way we either have to correct all occurrences of nullptr_t to
std::nullptr_t, or do something similar with providing nil under golang:: .

To reduce noise I prefer the later and let it be named as Nil.

The code was assigning nil to local, _not_ global _tblockforever. As a
result _tblockforever was left set with a test hook even after leaving
test context. Fix it.

The bug was there starting from 3b241983 (Port/move channels to
C/C++/Pyx).

Had to change `= nil` to `= NULL` because with nil Cython complains as

        def __exit__(pypanicWhenBlocked t, typ, val, tb):
            global _tblockforever
            _tblockforever = nil
                            ^
    ------------------------------------------------------------

    golang/_golang_test.pyx:86:25: Cannot assign type 'nullptr_t' to 'void (*)(void) nogil'

This is https://github.com/cython/cython/issues/3314.

It's a leftover originating from b073f6df (time: Move/Port timers to
C++/Pyx nogil).

Instead of `pyctx.ctx = nil` it was just `ctx = nil` - i.e. assign nil
to local variable instead of changing pyctx instance data. We were not
observing this bug because Cython, for C++ fields of cdef classes,
automatically emits in-place destructor calls in generated __dealloc__

    https://github.com/cython/cython/blob/0.29.14-11-g8c620c388/Cython/Compiler/ModuleNode.py#L1477-L1478

and so this way there was no leak. However we want to be explicit and the
code was not correct. Fix it.

The bug was there from 2a359791 (context: Move/Port context package to
C++/Pyx nogil).

Convert Pyx part of the project to use nil instead of NULL.

Not every usage of NULL was converted and some places were left to use
NULL where changing it to nil currently hits Cython compilation error:

https://github.com/cython/cython/issues/3314

Convert C++ part of the project to use nil instead of NULL/nullptr.

We do not convert pyx part yet, because Cython currently does not
understand that nullptr_t has properties of NULL and with e.g. the
following change

    --- a/golang/_context.pyx
    +++ b/golang/_context.pyx
    @@ -116,7 +116,7 @@ cdef cppclass _PyValue (_interface, gobject) nogil:
         __dealloc__():
             with gil:
                 obj = <object>this.pyobj
    -            this.pyobj = NULL
    +            this.pyobj = nil
                 Py_DECREF(obj)

errors as

    Error compiling Cython file:
    ------------------------------------------------------------
    ...
            if __decref():
                del self
        __dealloc__():
            with gil:
                obj = <object>this.pyobj
                this.pyobj = nil
                            ^
    ------------------------------------------------------------

    golang/_context.pyx:119:25: Cannot assign type 'nullptr_t' to 'PyObject *'

https://github.com/cython/cython/issues/3314

Nil is more native to Go.

https://godoc.org/lab.nexedi.com/kirr/go123/xsync#WorkGroup
kirr/go123@515a6d14