Python 3 compatibility via Cython

[1fish2]

• Rewrite the python-to-C glue obsidianmodule.c in Cython as arrowhead.pyx.
  • Cython is somewhat tricky, combining the rules of both Python and C, but it's much simpler, more readable, and more adaptable than the Python/Numpy C extension API. Cython diligently handles type checking, ref counting, and lots of other knotty details.
  • In a few test_arrow.py measurements, it ran about 5% slower, but that's within measurement noise. The obsidian.c code is unchanged except the caller passes in a struct instead of 13 arguments (to save some packing and unpacking code) so the solver should run at the same speed.
  • This almost avoided numpy's C API entirely. E.g. memoryviews are fast and easy way to access numpy (and other) arrays. But of the half dozen ways to copy a C array into a numpy array, numpy's C API generated significantly smaller object code. Still, I would've forgone that to avoid the np.import_array() rabbit hole if I'd only known. The fix is easy; knowing about it was not.
  • That problem was amplified by setup.py hell. Short answer: The current file seems to work reliably if you first delete the previous outputs.
• Use six to support some Python 3 renames.
• Update requirements and the stochastic-arrow version number.
• Fix some C compiler warnings (like printf format) and Python inspection warnings (like local variable time shadowing module import time).
• Some cleanups and README additions.
  • [It might be easier to read this PR one changelist at a time.]

TODO: Test building and installing pip packages into Python 2 and 3.

[1fish2]

Fixes #40

[1fish2]

Thanks for reviewing all this. It was a good opportunity to discuss Cython. Happy to do more of that.

BTW, I started a wiki page on Cython. So far it just links to key web pages. Maybe some of the points here could go there.

Also BTW, the Cython docs got significantly better in the last release.

I'll experiment a bit more with setup.py then merge this PR.

[1fish2]

Yes, I did test in python 3.