Potential use of CrabNet Hyperparameter benchmark task

[sgbaird]

It would be compelling to me to see the benchmark run on the CrabNet Hyperparameter optimization benchmark task (original work), which is a Turing test-like benchmark task I created that captures a number of advanced optimization problem characteristics (high dimensionality, heteroskedastic noise, multiple objectives) and seeks to be nearly indistinguishable from the original iteration (on average it took ~3 minutes each for the 179k runs on an RTX-2080 Ti GPU, with a total of ~1 year of GPU days consumed to create the benchmark).

See https://huggingface.co/spaces/AccelerationConsortium/crabnet-hyperparameter, which can be called programmatically (avoids having to host the large ~6 GB model on a laptop), as shown in https://github.com/AccelerationConsortium/baybe-multi-task-bo/blob/main/notebooks/HF-API-Ax.ipynb or https://github.com/AccelerationConsortium/baybe-multi-task-bo/blob/main/notebooks/3.1.2-AdvOpt-Ax.ipynb.