Performance assessment of creating geometries when partially evaluating a model (in the context of Gibbs sampling)

[amal-ghamdi]

Just a reminder from our review discussion: the geometries are repeatedly created, and this procedure might be simplified in the future

Originally posted by @chaozg in #664 (comment)