Add some state space model examples ported from dynamax

[theorashid]

Adds a "State Space Models" section adapted from some dynamax examples and reimplemented with the dynestyx API. Some changes from dynamax:

• Filters/smoothers use the cuthbert backend; ekf_mlp uses cd_dynamax for its Jacobian-based EKF (cuthbert's Taylor EKF doesn't fit the MLP)
• Dropped EM in lgssm_learning (dynestyx has no EM); Bayesian route uses NUTS
• ekf_spiral is EKF-only
• MLP rewritten with flax.nnx
• mkdocs.yml: new nav section + MathJax switched to SVG (fixes matrix delimiter rendering site-wide)

ekf_mlp emits a harmless cd_dynamax UserWarning

UserWarning: cd_dynamax discrete-time filters ignore absolute time arguments in GaussianStateEvolution/GaussianObservation. For genuinely time-varying discrete-time models, use a filter/backend that preserves absolute time semantics, such as EnKFConfig(filter_source='cuthbert').
  params_nl = gaussian_to_nlgssm_params(dynamics)

Don't think this is relevant here?

cc @juanitorduz: this should help with #264 to see how to do basic SSMs. We should port over an example for a basic ETS model (or similar structural model). At some point, I'd like to do an example like the online MLP where we get a new data point (in time) and we online update both the filtered forecast and the hyperparameter estimate. I know they're looking into a similar (non-forecasting) example in cuthbert for this too.

[DanWaxman]

Thanks Theo!! This looks great! I'll take a closer look after #266 merges (and try to take care of any merge conflicts that introduces).

UserWarning: cd_dynamax discrete-time filters ignore absolute time arguments in GaussianStateEvolution/GaussianObservation. For genuinely time-varying discrete-time models, use a filter/backend that preserves absolute time semantics, such as EnKFConfig(filter_source='cuthbert'). params_nl = gaussian_to_nlgssm_params(dynamics)

Don't think this is relevant here?

I agree

[juanitorduz]

Thanks @theorashid ! Let me take a detailed look 🙏 ! Exiting!