BO uncertainty constraints and other tweaks for safety-critical scenarios

[nikitakuklev]

This is the PR implementing new techniques used in recent FNAL Booster studies to ensure safety-critical pessimistic exploration. It adds constraints based on GP uncertainty, effectively penalizing regions that lack nearby samples.

Depending on how it is combined with noise priors, noise limits, and other fitting configs, the behavior of these constraints changes drastically from no impact to a very strong enforcement of local sampling. Care must be taken to ensure inferred and experimental noise levels are reasonable relative to constraint levels. We should expose the advanced config options but set defaults to something sane, like the 'untransformed absolute noise level'. This config can be worded as 'only sample where objective uncertainty < +-x', and has intuitive experimental interpretation.

While in theory universally compatible just like regular constraints, for some acquisition functions a special implementation that reuses computations can be faster (at cost of extra code).

This PR will provide core implementation, a debug notebook, and some visualization helpers. For now, config will be stored as generator field and not in VOCS, because only a few methods can provide uncertainty quantification.

[roussel-ryan]

Looks cool, when you have this ready we should schedule a discussion to see how this synergizes with #341