Bayesian recommendations from prior

[AdrianSosic]

This PR enables the use of Bayesian recommenders for cases where no training data is available (i.e., using the prior only). This is done by:

• ensuring that all dataframes are intact on the BayBE side even when they are empty
• making sure that the models still return the desired result when presented an empty training data set

[AdrianSosic]

Hi @AVHopp, @Scienfitz, this is the start of a PR enabling the use of the Bayesian recommenders without data, i.e. the missing piece to complete the TL example (#118) that allows to use the same recommender for the cases w/o training data, aligning the start of the learning curves. Unfortunately, due to time limitations, I could not manage to finish it, so either you take care of it or I do it after my return. The first part is done (i.e., ensuring that all dataframes are intact even when they contain no data) but now the botorch part needs adjustments.

[Scienfitz]

@AdrianSosic since the surrogate refactoring the way should be open to finish this now. Or are there further issues from the Botorch side (if so lets track them here)?

[AdrianSosic]

Just a quick comment to track the current state. At the moment, progress is (potentially?) still blocked by:

• [Bug] MultiTaskGP without data meta-pytorch/botorch#2360
• [Bug] Inconsistent behavior: `standardize` vs. `Standardize` with n < 2 meta-pytorch/botorch#2422

Requires another implementation attempt to see where exactly things break and how to avoid the problems.