Implement iterative NUTS algorithm

[ColCarroll]

This seems more readable, and perhaps modestly more efficient than the recursive implementation. As far as I can tell, it is also novel: the original c++ and matlab versions, as well as versions in julia, python, and R are all recursive (and use the original notation).

It passes all existing tests (except this test_parallel_start one?), but would appreciate another look.

[springcoil]

The changes look good to me after a first pass

[kyleabeauchamp]

It would be great to have a PDF version written up somewhere, as the recursive implementation has been a "barrier to entry" for myself and several colleagues.

[ColCarroll]

That failure looks funny, and might have more to do with conda than with anything else. It passed on my branch and when I reinstalled things locally -- I'll see if force pushing my branch will kick off another build that fixes it.

@kyleabeauchamp -- I'll type up the notes I made this weekend, which might also help to find any problems!

[ColCarroll]

apparently it is a thing! i'll look into it -- we should probably be pinning all the versions in conda.

[twiecki]

These conda version issues are pretty annoying. Maybe this is now fixed already on their end?

[springcoil]

Wow this error is annoying :)

[coveralls]

Coverage increased (+0.006%) to 81.096% when pulling 04bd9a6 on ColCarroll:iterative_nuts into 8e34b01 on pymc-devs:master.

[springcoil]

Ok I'll merge this now - great work @ColCarroll I look forward to that IPyNB or PDF write up or blog post (however you wish to do it) on how this algorithm actually works. I've not seen it anywhere else.