Allow for a GP to model covariance matrix

[richteague]

Posteriors are wildly underestimated, likely due to the poor (non-existent) handling of uncertainties. Try and implement a GP to model the covariance matrix and then marginalize over its hyper-parameters to see if that improves the returned posteriors.

[richteague]

This is probably the way to go: https://tinygp.readthedocs.io/en/stable/tutorials/multivariate.html.