Best practice for parallel model selection, especially avoidance of recompilation

@Dekermanjian

Hi. Thak you for sharing your approach.
Yes, I have tried it and I confirmed it works if we just change hyperparameters of the prior.

However this compiled model is not picklable and I couldn’t use it with joblib.Parallel. Perhapse I can try it with threadpool in the latest non-GIL Python 3.13.

On compilation time, for example this simple NB regression model took about 40 seconds to be compiled with nutpie+numba in google colab environment.