CI setup for manylinux wheels on tagged commit

Avoid raising a SystemError when clearing slots if setstate() failed.

PR #52 introduced a code path to `ghostify` that calls PyErr_Clear()
with the intent to avoid propagating AttributeErrors for slots.

However, if there is an error (like a POSKeyError) raised by
jar.setstate(), then `unghostify` will call ghostify with an error
pending. If the object had slots that weren't set and the
AttributeError was cleared, so was the pending error from setstate. So
when `ghostify` returned NULL that got propagated up to the
interpreter which finds no exception and so raises `SystemError: error
return without exception set`.

This commit makes `unghostify` save and restore the exception state
around the call to PyErr_Clear.

Fix ``__setstate__`` interning when ``state`` parameter is not a built-in dict

Use macpython for all Mac builds.

Ref #58

Python 3.6 Mac

Support Python 3.6

new_ghost shouldn't clear newargs (PyPy consistent with C)

Expose PersistentList/Mapping docs. Fixes #52.

PickleCache.

Fixes #49.

Unfortunately, this doesn't fix the whole problem, because
_p_deactivate() still clears them, and so we lose what we set in __new__.

@jimfulton and I have talked about it, and I'm (mostly :) convinced that
this shouldn't be an actual problem for any of the reasons described in
the previous comment.

If I use PyPy 5.4.1 to run the ZODB master test suite against
this (well, with #44 rolled back) I don't get any unexpected
failures. (I haven't run the ZEO test suite yet.) Which honestly amazes
me because I'm sure I used to get test failures---I guess the PyPy GC
has changed...which means we may see some failures on Travis. I'll try
to set up an older PyPy to verify.

Fix Python timestamp hash on 64-bit windows

This reverts commit 358cd618.

The _wraparound function is still broken on PyPy3 (both the old version
on Travis and the latest available) if we don't use ctypes.

Fixes #51

Part of the problem was in the tests. The detection of when the hash
gets truncated was broken on win32.

Part of the problem is somewhere in the _wraparound function, I think,
because even fixing the tests to use the right values we still get
mismatches. Rather than dig into that, I just go back to using ctypes
when it's available.