Major performance improvements: Ensembles, Volumes, Splitting - #455

Performance improvements. Five-fold speedup in the DNA analysis I was doing; probably less (but still significant) in our examples.

• Move expensive debug logging into logger.isEnabledFor(logging.DEBUG) scopes
• Add short-circuit logic to volumes
• Fix problem that .split is using an algorithm that does not scale linearly (not using trusted=True)
• Switch ensembles to be better for short-circuit (Length before Volume) in flux calculation

Timings (in order of my implementations, so these are cumulative improvements) to use SingleTrajectoryAnalysis.analyze_flux() on a DNA trajectory of 10k frames. Inexact, since I just ran the test once on my laptop, but when you see a calculation drop from ~5 minutes to ~1 minute, that’s well beyond the margin of error (even for my sample of $N=1$).

• Current master branch: 264.98s user 1.95s system 89% cpu 4:56.82 total
• Rescope debugging: 127.29s user 1.64s system 89% cpu 2:24.61 total
• Short-circuit volumes: 96.98s user 1.45s system 90% cpu 1:49.07 total
• Fix split algorithm: 62.83s user 1.35s system 80% cpu 1:19.98 total
• Fix short-circuit order in flux: 57.13s user 1.20s system 85% cpu 1:08.24 total

Of course, getting such improvement from the short-circuit volumes requires defining your volumes to take the most advantage of short-circuit logic. But that was easy to do with this particular system.

There’s some room to improve SequentialEnsemble.can_append, especially in its use of _find_subtraj_final. This should give us a significant speed boost. I’ll do that at a later date.

Other notes for future speedups:

• Looks like calculating the unit cells (after loading trajectory) is one of the major bottlenecks for this (25% of the total time). But I think that is in MDTraj?
• chaindict is always going to be inner-loop stuff. It looks like a good target for cythonization -- it does a lot of python function calls deep in inner loops, and Cython can massively decrease the overhead cost of those function calls.
• There are a number of algorithmic speed-ups that can happen once we start including results from path ensemble theory. That should be after 1.0, though.