Investigate Removing .params

[ilan-gold]

We should look into removing params as it is causing issues and is not necessary. If one needs the concatenation of the fitted parameters, then that should be a feature and not the default.

Specifically, the ModelContainer has setters for theta_location theta_scale that are quite slow because of having to update one .params dask array which is then recreated.

This should allow for theta_scale and theta_location to be optional (i.e for Poisson models)