Support inequality constraints for knowledge gradient as long as they are not inter-point

[esantorella]

See pytorch/botorch#1220 for context. BoTorch does support inequality constraints (and probably also equality constraints) with knowledge gradient, as long as they are not inter-point. That means that when the indices field of a constraint is 1d, it's okay, but when it's 2d, there will be an error.

Currently, Ax will error if any inequality constraints are provided, even though constraints that are not inter-point will work. So this error can become

if inequality_constraints is not None:
    indices, _, __ = inequality_constraints
    if indices.ndim >= 2:
            raise UnsupportedError(
                "Inter-point constraints are not yet supported for KnowledgeGradient!"
            )

Then there will be some further work needed to use the constraints in the 1d case.

[lena-kashtelyan]

Merged to a larger KG issue.

[esantorella]

This should be supported by using botorch_modular with inequality constraints.