Review Connection/DB closure

In addition to some micro-optimisation, this fixes the following minor issues:
- Closing a DB left Connections registered to the global transaction manager.
  Which broke at least multi-db because only such primary connections were
  ignored with monkey-patches.
- Stop calling the real newTransaction when aborting an existing transaction.
  This was not the case if explicit_transactions=0.

Unpin specific PyPy versions, blacklist 3.3 in `setup.py`

* Use a higher pickle protocol (2) for serializing objects on Python 2

Previously protocol 1 was used. This is more efficient for new-style
classes (all persistent objects are new-style), according to the docs,
at the cost of being very slightly less space efficient for old-style
classes.

In tests of a persistent object with two trivial numeric attributes,
the higher protocol was 12 bytes smaller, and serialized and
deserialized 1us faster. Introducing a reference to another new-style
class for a more realistic test made the higher protocol twice as fast
to serialize (20.5 vs 10.3us), almost half the size (215 vs 142
bytes), and it deserialized 30% faster (6.5 vs 4.6us).

On Python 2, this will now allow open ``file`` objects to be
pickled (loading the object will result in a closed file); previously
this would result in a ``TypeError`` (as does under Python 3). We had
tests that you couldn't do that with a BlobFile so I had to update it
to still make that true.

I wouldn't recommend serializing arbitrary open files under Python
2 (for one thing, they can't trivially be deserialized in Python 3),
but I didn't take any steps to prevent it either. Since this hasn't
been possible, there shouldn't be code in the wild that is trying to
do it---and it wouldn't be forward compatible with Python 3 either.

[ci skip]

Add badges to README.

[ci skip]

HistoricalStorageAdapter forwards release() to its instance

[skip ci]

Fixes #78.

Also adjust the base wrapper's __getattr__ to be faster (500 vs 711ns
in a benchmark of the common case) and behave the same on Python 2 vs
Python 3 with regards to exceptions, should a method be implemented as
an descriptor. In practice this probably makes no difference
whatsoever, it's just the same pattern used for forwarding
release.

Add support for 3.6, drop 3.3

[skip appveyor]

Remove outdated comment.

Remove redundant usedevelop.

[skip ci]

We need a newer setuptools on older versions of Python to be able to install recent Sphinx because of its use of environment markers.

Also add note on supported versions to the docs and readme.
Fixes #168.

Update tox.ini to correctly run testdocumentation and DRY with test deps.

* clean up references to DBs and Connections in pools
* Implement an explicit clear() method for clearing all the items in the pool.

* Make all classes new-style.

On PyPy, it's documented (http://pypy.org/performance.html#id3) that
old-style classes are slow, and classes that derive from both old and
new are especially slow. I checked with the PyPy devs on IRC today and
they confirmed that's still true with the latest PyPy2 releases.

Unfortunately, FileStorage was one such mixed class, as was Connection.

Moving to new-style classes seems to have a positive impact in the
benchmarks. Here's zodbshootout on PyPy2 5.7.1 against a
FileStorage (running in --threads and with 5 reps to be sure we get
warmed up). First, current ZODB:

"Transaction",                fs
"Add 1000 Objects",            31,738
"Update 1000 Objects",         42,444
"Read 1000 Cold Objects",      51,894
"Read 1000 Hot Objects",       53,187
"Read 1000 Steamin' Objects", 835,877

And with this PR:

"Transaction",                fs
"Add 1000 Objects",             35,651
"Update 1000 Objects",          54,906
"Read 1000 Cold Objects",      103,484
"Read 1000 Hot Objects",        84,721
"Read 1000 Steamin' Objects", 2,112,095

The tests that hit the storage extensively are notably faster, as are
steamin and hot, Connection having been a mixed class.

I ran all tests multiple times. The data files were removed between
runs. There's some variation, but the new-style classes always seem
better.

For comparison, here's CPython 2.7.13:

"Transaction",                fs
"Add 1000 Objects",            19,531
"Update 1000 Objects",         16,201
"Read 1000 Cold Objects",      22,111
"Read 1000 Hot Objects",       21,851
"Read 1000 Steamin' Objects", 880,582

Locally I've run the tests under 2.7 and they all passed.

Profiling (https://github.com/zodb/zodbshootout/pull/32/) showed that
this method was the only blob-related method that showed up in a test
of creating blobs, other than those that actually performed IO.

With this change its total and cumulative time drops from 0.003/0.004
to 0.001/0.002 in a small benchmark. Blobs created per second shows a
small but consistent improvement.

Before:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
      100    0.005    0.000    0.005    0.000 {built-in method rename}
      100    0.004    0.000    0.004    0.000 {function BlobFile.close at 0x1080d3a60}
      200    0.003    0.000    0.004    0.000 blob.py:576(oid_to_path)
      101    0.003    0.000    0.003    0.000 {built-in method mkdir}
      100    0.002    0.000    0.002    0.000 blob.py:333(__init__)
      402    0.002    0.000    0.005    0.000 {method 'dump' of '_pickle.Pickler' objects}
        1    0.002    0.002    0.034    0.034 Connection.py:553(_store_objects)
      201    0.002    0.000    0.002    0.000 {built-in method stat}
     5633    0.001    0.000    0.002    0.000 {built-in method isinstance}

After:

   ncalls  tottime  percall  cumtime  percall filename:lineno(function)
      100    0.005    0.000    0.005    0.000 {built-in method rename}
      101    0.005    0.000    0.005    0.000 {built-in method mkdir}
      100    0.004    0.000    0.004    0.000 {function BlobFile.close at 0x10636aa60}
      402    0.002    0.000    0.005    0.000 {method 'dump' of '_pickle.Pickler' objects}
      100    0.002    0.000    0.002    0.000 blob.py:333(__init__)
        1    0.002    0.002    0.035    0.035 Connection.py:553(_store_objects)
      201    0.002    0.000    0.002    0.000 {built-in method stat}
     4033    0.001    0.000    0.001    0.000 {built-in method isinstance}
   ....
      200    0.001    0.000    0.002    0.000 blob.py:576(oid_to_path)

Fixes #158.

And emacs backups. :) Because of the way zest.releaser works, this
isn't strictly necessary, but it makes the manifest checker shut up.