Are there instances when reparameterizing a model makes it slower?

How many groups are there, just two? Because two is too few to use a hierarchical distribution.

My guess is the centered model is running slowly because all_groups_mean is given low prior likelihood of a value that extreme (-4).

Try modeling as an intercept + dummy for treatment, and make set the prior for the intercept to be something like pm.Normal('intercept', 0, 4)