Dict-based objects are easy, just an assignment. But objects with
__slots__ require generating new classes on-the-fly with matching
descriptors (which is how slots are actually implemented). One
prominent use-case is the `persistent` library which is slot-based.

This is only implemented for the pure-python version because it's only
been observed to be a problem in that use-case.

Fix #7. Make the Python Acquirer cooperatively call super, and test this. Also provides a more informative traceback if lookup failed in that case which assists debugging if the inheritance hierarchy is complex.

When temporarily decoding an incoming attribute name from a unicode string, be careful not to destroy the temoporary PyObject until we're done using its char* buffer. Fixes zopefoundation/Acquisition#5.

Set PURE_PYTHON=1 for coverage testing under tox.

A Python-implementation-specific test for the repeated-search-optimization gets us to 100% coverage.

Also fixes a few discovered differences between the C and Python implementations.

Add a test and a fix for the __of__ issue @davisagli reported. Only tested under Python2 because I couldn't figure out how to reproduce it under Python 3

Adapt the garbage collection tests to run under PyPy. This makes the overall test diff smaller and easier to read.

Fix the last doctest errors in pure-python mode about parent circles. They were another manifestation of zopefoundation/ExtensionClass#3 and required a workaround.

Add tests for all the methods proxied by the C wrappers, and make the Python implementation match that behaviour.

All tests continue to pass under 2.6 and 2.7.

The pure-python code currently raises AttributeError instead of RuntimeError in the
recursive-parent cases, though I'm not sure why that is. There may also be some issues
with the proxying of some methods, but those aren't covered by current test cases.

setup.py: Set url to github

See: http://blog.travis-ci.com/2014-12-17-faster-builds-with-container-based-infrastructure/