Add general timeseries support for `joint_logprob`

[brandonwillard]

We can add support for Scan Ops so that joint_logprob would cover nearly arbitrary timeseries.

A lot of the relevant Scan rewrite code was written for symbolic-pymc in pymc-devs/symbolic-pymc#113 and pymc-devs/symbolic-pymc#114, so we can take some of the code/ideas from there and adapt it to aeppl.

See the Gitter conversation about this idea here. (NB: If you use https://develop.element.io/ you can also turn on LaTeX rendering in the "Labs" settings.)

Also, see these tests for explicit examples of log-likelihood generation for a simple DLM and HMM in symbolic-pymc.

[ricardoV94]

Also with a cumsum we can cover many vanilla RandomWalks

[brandonwillard]

If we don't already, we should also have rewrites that turn cumsum Scans into actual cumsums.