Maintenance Release, MultiTaskGP Improvements

New Features

• New default parameterization for MultiTaskGP (#3049).
  See discussion: #3065

  • Add PositiveIndexKernel (#3047).

• Add HeterogeneousMTGP for transfer learning between different search spaces (#3073).

Bug Fixes

• Add missing int cast in MCSampler (#3062).
• Fix shape error in qNegIntegratedPosteriorVariance (#3068).
• Add taus to the model state dict for EnsembleMapSaasSingleTaskGP for consistency when loading state dict (#3069).
• Fix confusing warning when using input transforms with deterministic models (#3078).

Other Changes

• Structure options arguments in gen_candidates_torch (#3019).
• Set default cache_root based on whether the model supports it (#3075).
• Use nanmean and nanstd in Standardize outcome transform (#3072).