Performance issues

[ColCarroll]

As we grow the example gallery, we should probably also check out performance compared to a direct implementation in TFP, and make sure PyMC4 is not adding too much (any?) overhead, or that we at least understand where the overhead is coming from.

I will followup with a reproducible example, but the partial pooling example from PyMC3 runs more than 10x slower (a few minutes instead of a few seconds) on PyMC4 compared to TFP.

This can be a useful thread for collecting such examples.