Currently, trying to compile (any?) statespace model in mode="NUMBA" fails. It would be really, really nice to fix this.

I think the error is related to advanced indexing here. This code is responsible for masking out any missing values at a given time step (they are treated as hidden states for that update). It shouldn't be hard to write this check in a different way that numba will be happy with. I'm not 100% sure that's the actual problem, but it's where I'd start.

We should make a test_statespace_NUMBA.py file that more-or-less copies the test_statespace_JAX.py file, then do a red-green refactor on the code. Actually, we should probably start with a test_kalman_filter_NUMBA file, since that will help break the problem down into smaller pieces and identify the bug. Testing the filters alone (without any of the complexity overhead of the models) would be good.