Consider re-writing innovation-based parameterizations of key stochastic processes as `numpyro.Distributions`

[dylanhmorris]

Avoiding the scan in log prob evaluation and using it only in forward sampling is a really nice feature of doing this as true numpyro Distributions, rather than as PyRenew RandomVariables.

Question: does this argue for reimplementing the innovation-based parameterizations in the same way? Do you expect performance implications?

Originally posted by @dylanhmorris in #828