Rename `gp.predict` to `gp.compile_predict`

[ricardoV94]

Description

Anything that compiles a Pytensor function under the hood and does a single call should be changed to returning the compiled function and allowing users to call it as many times they want.

A user on the office hours was asking how to speedup predict.

This is similar to the switch we did for logp->compile_logp

Also the function seems to be reinventing some of the functionality of PointFunc. Can't we use model.compile_fn under the hood?