This makes a number of improvements to sequential_gaussian_filter_sample():

1. Allows user to pass in noise, rather than drawing noise internally (i.e. more JAX-style). The motivation here is to allow both (1) computing the mean, and (2) supporting antithetic sampling where we can pass in cat([noise, -noise]).
2. Avoid dropping the first entry of the returned samples. This makes sequential_gaussian_filter_sample() more useful outside of the context of GaussianHMM, e.g. it will be easier to use in a fully-observed Markov model. This attempts to make up for my admittedly bad design choice of making the hidden Z series one step longer than the observed X series in GaussianHMM 😬
3. Adds a profiling script. I saw no speed changes due to this PR.
4. Refactors some backward-sample logic to use placement into a torch.empty(), rather than torch.nn.functional.pad and torch.stack. This is admittedly less functional, but does reduce memory usage and IMHO reads cleaner.

This also exposes the helper matrix_and_gaussian_to_gaussian() which I'm finding useful.

Tested

• added new gradient tests
• added tests for antithetic sampling
• profiled with and without gradients, saw no time difference