Fix #21, FileStorage: when disk space runs out while packing, do not leave the ``.pack`` file around

Support PyPy. 

This is a performance optimization for PyPy (profiling shows many recursive calls to Persistent.__getattribute__, and recursive calls seem to trouble the PyPy JIT; this eliminates many of them when loading objects):

"Transaction",                 AFTER       BEFORE
"Add 3000 Objects",             20205      18505
"Update 3000 Objects",          24191      18807
"Read 3000 Warm Objects",       28920      26671
"Read 3000 Cold Objects",       28745      27487
"Read 3000 Hot Objects",        29563      27487
"Read 3000 Steamin' Objects", 1033758     916587

Go back to assuming that cache.lru_items and cache.cache_data are in sync. Otherwise we have corruption in the cache and, under PyPy, a CFFI-level memory leak; we want tests to fail in that case.

Fix test_new_ghost_w_persistent_class under PyPy3, failing due to a bytes/unicode issue. It seems the cPickleCache under Python 3 accepts unicode objects as the oid, but the pure-Python version doesn't.

Rollback some changes now that Python persistent handles registering objects during __setstate__ the same way as C persistent.

Fix unpickling POSError subclasses under PyPy. This fixes a bunch of '<Unprintable POSKeyError>' tests in ZEO.

The Python persistent implementation has caught up enough with the C version that we don't need to list extra special names that shouldn't re-wake Broken objects.

Fix the functional blob tests under PyPy/pure-python by avoiding spuriously registering Blob objects as modified each time they are loaded from cache. This is something of a difference in behaviour between the C and Python implementations, but it's also arguably a bug in the Blob class.

The changes were mostly in the Persistent module. Here, the changes were minimal:

- Introduce zodbpickle for PyPy and Python 2.7 to fix noload.

- Centralize the construction of Picklers/Unpicklers to account for the differences between Python2/3/zodbpickle.

- A few extra gc.collect() calls.

- Some minor printing differences in the doctests due to the Python implementation of BTrees.

Checkpoint on ZODB under PyPy: 25 failures, 25 errors (needs my persistent fork). Many of them are printing differences, some have to do with the Python persistent implementation, especially PickleCache

Use zodbpickle under pypy and 2.7 for a working noload. Also add pypy to tox so I can start testing it.

Fix manifest and setup

- include .coveragerc
- Add author, author_email, keywords

    $ check-manifest
    lists of files in version control and sdist match
    $ pyroma .
    ------------------------------
    Checking .
    Found ZODB
    ------------------------------
    Final rating: 10/10
    Your cheese is so fresh most people think it's a cream: Mascarpone
    ------------------------------

Add tests for repozo's parseargs().

Fixes https://github.com/zopefoundation/ZODB/issues/28

It didn't work for me, because nose would crawl inside manual_tests, try
to run test_fstest.py, and fail (because that test is manual and
requires extra setup, like putting fstest.py on PYTHONPATH).

I also don't think that we want to exclude test layers from coverage
measurements (nose doesn't support them).

This change makes 'tox -e coverage' use zope.testrunner like it does for
regular tests.

See https://github.com/zopefoundation/zope.testrunner/pull/19

Cleaned up exception logging and changed pins