Possible 25x speedup in steady-state `inner_loop` function

[nikhilwoodruff]

I've found that eliminating dask multiprocessing can speed up the inner_loop function's main loop by around 25x (each iteration: ~0.75s without dask, ~20s with dask on my machine), probably due to high startup times. This might not apply to every machine.

Now, I can fail to find the steady-state 25x faster! 😄

@rickecon @jdebacker @jpycroft @MaxGhenis