Cache the number of elements in the action space

[thatguy11325]

You probably dont need to dispatch to numpy everytime you call split to calculate the number of elements in the space. This PR caches the sizes (in a less than nice way imo) as an example. Before and after pictures below

[jsuarez5341]

This looks reasonable, waiting for hardware to test end to end. Any other optimization ideas for split? It's the main bottleneck right now. From before your patch:

🐡 python tests/test_extensions.py
0.00000032: Flatten time
0.00000294: Concatenate time
0.00001958: Split time
0.00000056: Unflatten time

[thatguy11325]

You could try to vectorize the generation of samps -> leaves a la what was done in evaluate? Though I'm unsure if that'll work if the sz's can vary.

[thatguy11325]

I think it'd look like

leaves = stacked_sample.reshape(len(flat_space), batch, *next(flat_space.values()).shape)

I assume since sz is the same across all flat spaces, shape will be the same too.