Class level dimension and parameter information

[lucianopaz]

I opened this issue just to keep track of this pymc3 discussion. Whatever design choice we end up taking there, should also be addressed here in pymc4.

[brandonwillard]

If I recall correctly, TFP uses effectively the same shape/dimension design outlined here and later implemented in the Theano RandomVariable, but under a different naming scheme (e.g. "batches" or "event" instead of "replications" and/or the "independent" dimensions I used). This means we might already have the information needed for 1); however, I'm not sure about 2).