[Another] Numba acceleration, replacing cython with automatic parallelism

[FractalMachinist]

To address limited performance, two main changes:

1. Don't wait for the drawing to finish

Move the matplotlib drawing step to a separate thread. This is responsible for a very significant speedup, but increases complexity (more in Pros/Cons). The single largest contributor to the drawing step is the flow lines, which can't be accelerated without a very significant re-write of matplotlib. This does mean we can compute far more than one step per drawing, which invites higher-resolution models.

2. Let Numba manage parallelism

Instead of writing explicitly multi-threaded model evaluation, use the Numba compiler (getting rid of Cython) in ways that allow for automatically managed multi-core execution.

Pros:

1. Automatic parallelism and optimization without asking for dual fluency in Python and Cython
2. Significantly increased execution speed at every step
3. Headroom for higher-resolution models

Cons:

1. The matplotlib drawing thread.
   • All plotting data has to get explicitly passed through a multiprocessing Queue.
   • These changes are invisible unless the drawing system needs modified (EG it's build-and-forget).
2. Any common multiprocessing library is going to require knowledge beyond base python, and while Numba is >95% original Python, it's still more complex than no multiprocessing.

[arenmarsden]

@FractalMachinist I'd remove commit 221c8d4

[TargetPolymath]

@arenmarsden Oops, you're absolutely right.

EDIT: I don't know how to edit this particular pull request in-place. I'll get the Doctor's attention during the stream if it comes up.

[ATGSilva]

Stale