Fix kernel translation

[AdrianSosic]

This PR fixes one of the two issues that caused transfer learning performance to deteriorate, fixing several other issues along the way.

Bug fixes

• Main issue fixed: the arguments provided to to_gpytorch (e.g. ard_num_dims) were not passed on to the GPyTorch kernel class.
• The to_gpytorch call was not tested and did in fact not succeed in all cases. Causing issues:
  • Hypothesis strategies did not account for positive-valued attributes, and no validators where in place. --> fixed
  • The torch dtype was not correctly set when translating priors to GPyTorch. --> fixed using temporary workaround (see below)

Added workarounds

• A proper solution to setting the default torch dtype requires overriding the torch import mechanism using importlibs machinery, due to lazy loading. A preliminary workaround is added in form of the set_default_torch_dtype utility, which manages the temporary code in a single place.

Improvements

• The new match_attributes utility, which replaces filter_attributes now raises an error when attributes cannot be matched to the target callable, instead of silently dropping them. In strict=False mode, the unmatched arguments can be inspected as second return value for further processing.
• Kernel translation to GPyTorch now ensures that all attributes are actually mapped
• Added kernel validation, ensuring that invalid arguments trigger the expected errors
• Added kernel translation tests, ensuring that all kernel components are correctly mapped to corresponding GPyTorch components

[AdrianSosic]

Hi @AVHopp, @Scienfitz, here a few open issues we might want to discuss.

[AVHopp]

Minor comments. Will properly review in the following week. In any csae, kudos for spotting and fixing.

[AVHopp]

Looks very nice, have some questions about why some things are necessary/whether I understand them correctly.