Learned motion and sensor models

[hidmic]

So far all motion models and sensor models in Beluga are based on approximations derived from first principles and even practical heuristics. I think it's time to add support for data-driven models. We can use libtorch as the workhorse. Since most PFs already assume that the underlying system is Markov, we can backprop through one predict-update cycle. Beluga won't interfere with automatic differentiation. If we want to backprop across multiple steps, systematic resampling may help.

[hidmic]

I think this warrants a separate beluga_torch package, with provisions to help data collection and training.

[hidmic]

@glpuga this may be interesting at some point in the future for some F1/10 race car I know of.