Add BoTorch kernel preset, which uses dimensions-scaled prior

[Hrovatin]

This can stay a draft until it is validated that the performance improves with this preset and that we even want to use it.

Changes:

• Use base_covar_module that matches parameters from BoTorch at the time of writing.
• As long as this is separated from using MultiTaskGP this must be pre-defined rather than using defaults to enable the use of index kernel in SingleTaskGP. -> TODO after we switch to MultiTaskGP this PR should be adapted to simply use BoTorch default rather than re-implementing custom kernel factory.

Benchmarks:

• 4 cases as a grid of:
  • Whether using TL campaign with source data (TL) or using non-TL campaign with target data only (nonTL).
  • Optionally use Botorch preset (appended botorchPreset)
• The nonTL line should equal some of the lines from Add benchmarks #491 branch (compare to lines named 10 (10% of source data) or TL for benchmarks where only one source data size was used). However, that is not always the case. No idea why.

[AdrianSosic]

Hi @Hrovatin, just two comments upfront, to save us some work later:

• The point of the botorch preset should not be to cook up the same prior logic again in our code but to make sure that the GPs are simply called without any additional arguments. The noticable difference is that the latter automatically i) avoids any potential bugs from our end plus ii) will automatically adapt when botorch makes changes
• The preset should not only affect the kernel priors but all priors. This is also automatically fulfilled by not passing anything to the constructor.
• I suggest you split up the PR into two, one that brings the preset (which we want to have anyway, regardless of benchmark performance) and another that adds the benchmarks. Otherwise, this will be a ton of code to review

Cheers 🙃

[Hrovatin]

Ah then I strongly misunderstood the aim - we should then align during the meeting next week