This is more follow up to https://bugs.python.org/issue37250

The action taken is:
 - restore tp_print to not break all of the sdists on pypi for py38
 - remove tp_print for real in py39

In https://github.com/cython/cython/pull/3171 tp_print was initialized
for PY_VERSION_HEX >= 0x030800b4 however, when trying to use cython
with cpython master (aka py39) there are compile time exceptions due
to too many initializers:

   error: too many initializers for ‘PyTypeObject’ {aka ‘_typeobject’}

This fixes that by putting an upper bound on the ifdef of including
pt_print at the end of the object.

Make sure to include "Python.h" also from the "public" header file, which depends on its declarations.
Closes #3133.

Explicitly initialize tp_print in Python 3.8

When compiling cython-generated extension modules in Python3.8rc1
this error is emitted by the compiler:

  _ext.cpp:8104:1: error: missing initializer for member ‘_typeobject::tp_print’ [-Werror=missing-field-initializers]

The reason is that Python3.8 moved the tp_print slot (d917cfe4051) to
the end of the _typeobject struct and reused the original position for
tp_vectorcall_offset. The current generated code does not initialize the
deprecated tp_print slot that was moved to the end of the struct.

When compiling to C++ code and using automatic conversion from Python
str to C++ string, Cython inserts this function template but generates the warning:
warning: string.from_py:15:63: local variable 'length' might be referenced before assignment

* Fix coverage path search
* Add testcase

Issue: #2541

I was going to use posix_madvise(POSIX_MADV_DONTNEED) in my tests to
make sure a memory page should be evicted from kernel cache, but found
out that this operation is hardcoded to be NOOP on Glibc/Linux systems:

https://sourceware.org/git/?p=glibc.git;a=blob;f=sysdeps/unix/sysv/linux/posix_madvise.c;h=c89fa64f0749;hb=HEAD#l25
https://sourceware.org/git/?p=glibc.git;a=commitdiff;h=8889e7aa461a

On the other hand, as also verified by strace, `madvise(MADV_DONTNEED)`
is not a NOOP as it actually goes to kernel by making syscall.

Since it is not possible to use posix_madvise for what should be
POSIX_MADV_DONTNEED behaviour, let's add all raw madvise behaviour flags
as documented on http://man7.org/linux/man-pages/man2/madvise.2.html

Presumably MADV_NORMAL and other behaviours that should translate 1-1 to
POSIX ones were omitted in 59ac9991 (Add pxd for mmap & friends
from "sys/mman.h" covering POSIX/Linux/BSD) to make people use the
standardised POSIX interface. However given that it turned out to be not
possible to use posix_madvise for all standard behaviours,
system-specific madvise behaviours have to be used.

(cherry picked from commit b394295b)

PyGILState_STATE was added in commit 3fd6fdce (Gilnanny + pystate.pxd)
with struct type on-purpose different from what CPython actually uses
with the idea so that PyGILState_STATE cannot be coerced into int.

However as it is, it prevents PyGILState_STATE usage:

	cdef PyGILState_STATE gstate = PyGILState_Ensure()

gives

	Variable type 'PyGILState_STATE' is incomplete

Fix it by making PyGILState_STATE a defined structure.

(cherry picked from commit 82a0ed43)

mmap returns void* and indicates failure via MAP_FAILED that is defined
to be ((void *)-1). Every mmap call has to be checked for failure and
thus posix.mman with mmap, but without MAP_FAILED is not very useful.

We cannot declare MAP_FAILED as enum, because enum assumes int. However
Cython documentation says

	4. If the header file uses macros to define constants, translate
	   them into a normal external variable declaration. ...

	(https://cython.readthedocs.io/en/latest/src/userguide/external_C_code.html#referencing-c-header-files)

So add proper declaration for MAP_FAILED to make posix.mman.mmap usable.

(cherry picked from commit 67253049)

mlock2 is Linux-specific version of mlock that accepts flags. The only
flag available so far is MLOCK_ONFAULT which asks to lock pages only at
the time when they are faulted in, not in the beginning. MCL_ONFAULT is
Linux-specific flag to mlockall which requests similar behaviour.

http://man7.org/linux/man-pages/man2/mlock2.2.html

(cherry picked from commit 83a5700b)

Closes GH-3023.

Pythran assumes shapes are shapes of long, whatever the underlying
architecture.

Currently Cython generates code like this:

     int __Pyx_InitCachedBuiltins(void) {
         __pyx_builtin_NAME = __Pyx_GetBuiltinName(...);
         if (!__pyx_builtin_NAME) __PYX_ERR(1, 44, __pyx_L1_error)
     }

     int __pyx_pymod_exec_MODULE(PyObject *__pyx_pyinit_module) {
         if (__Pyx_InitCachedBuiltins() < 0) __PYX_ERR(1, 1, __pyx_L1_error)
     }

When InitCachedBuiltins and InitCachedConstants call __PYX_ERR, they
pass the file and line where a builtin is used, but then pymod_exec
overwrites it with 1 and 1, and the error message looks like this:

       File "FILE", line 1, in init MODULE.
         import os
     NameError: name 'NAME' is not defined

After this change Cython generates:

     int __pyx_pymod_exec_MODULE(PyObject *__pyx_pyinit_module) {
         if (__Pyx_InitCachedBuiltins() < 0) goto __pyx_L1_error;
     }

and prints:

       File "FILE", line 44, in init MODULE.
         print(NAME)
     NameError: name 'NAME' is not defined

Fix a C compiler warning about an implicitly reduced integer range (Py_ssize_t -> int) when calling a helper function.
Closes GH-3028.

Closes GH-3031

* REF: make Utils dependency-free by moving search_include_directories to Main
* BUG: prefer sys.path before Cython/Include when searching for pxd files