Performance bump in PSTH

[mmyros]

Single-threaded speedup through JIT histograms and parallelization over neurons. This introduces numba and joblib dependencies, but there is an option to turn off parallelization. If it's turned off, joblib is not imported. -1 option means use all available threads.

On Steinmetz Neuropix dataset, this gets a 3x single-threaded speed increase and theoretically linear speed increase with the number of threads

Tests are passing, except for strf: TypeError: object of type <class 'float'> cannot be safely interpreted as an integer. but I think that's unrelated since this PR does not deal with ML part. Edit: Fixed that test, indeed it wasn't related to this PR.

[mmyros]

Currently the failing test seems to do with tensorflow version. It's probably enough to pin TF to 2.2