Add Smoothers

[DanWaxman]

This PR adds smoothers under a new Smoother(...) handler, mirroring the current Filter(...) interface. It introduces a number of different smoothing algorithms from CD-Dynamax/cuthbert, some new documentation, and also introduces a bit of a refactor to integration code that was getting quite long (now split into filtering, smoothing, and util files).

Currently, prediction is not handled the way I would like it to be: I would like prediction to use the smoother within the smoothing window, and only pass forward the final predict_times as a filtered_times/filtered_dists pair. This is not currently done because missing data is not properly implemented in cd_dynamax for continuous time, and our indexing is not really correct for discrete time. For now, we decided to properly handle predict_times >= max(obs_times), and defer a more complete treatment to a later time.

• Include Smoother Handler
• Integration
  • Cuthbert
    • Particle Smoother
    • Kalman (RTS) Smoother
    • Extended Kalman (RTS) Smoother
  • CD-Dynamax
    • Kalman (RTS) Smoother
    • Extended Kalman (RTS) Smoother
  • Dynamax
    • Kalman (RTS) Smoother
    • Extended Kalman (RTS) Smoother
    • Unscented Kalman Smoother
• New tests
  • Smoke tests
  • Science tests
• Documentation
  • API Reference Pages
  • Full docstring & references for configs
  • Tutorial notebooks
• Proper prediction behavior

[DanWaxman]

@mattlevine22 I think this is ready for a look now!

[mattlevine22]

I'd add:

"Recall that Filter does, however, already support prediction times within the observation window."

[DanWaxman]

Done!

[mattlevine22]

Looks great!

1. Why expand Filter API to include record_smoothed_*?

2. Sharing functions between filters + smoothers:

• Seems like _cuthbert_states_to_dists, _posterior_to_dists, _particle_to_batched_dists can all be shared by filters + smoothers via very slight expansion (e.g. add a filter vs smooth mode to their arguments and very few additional lines)

[mattlevine22]

Approving, with recommendation to remove the unnecessarily thin wrappers in smoothers.py and filters.py , e.g. _smoothed_posterior_to_dists can just be removed by using _posterior_sequence_to_dists directly. This happens a few times in the filter/smoother files.