Refactor GPClassificationModel to use Botorch's InducingPointAllocator's

[crasanders]

GPs can make use of inducing points. Basically, instead of fitting the posterior to every training point, we can instead fit to to a smaller set of inducing points to make fitting more computationally efficient.

AEPsych implements some heuristics for selecting these inducing points in a utility function: https://github.com/facebookresearch/aepsych/blob/main/aepsych/models/utils.py#L54

Botorch implements some more principled algorithms as InducingPointAllocator classes: https://github.com/pytorch/botorch/blob/main/botorch/models/utils/inducing_point_allocators.py#L252
https://github.com/pytorch/botorch/blob/16853b46da337c8f9a4f51386d1c5eb690f337ad/botorch/models/approximate_gp.py#L168

AEPsych can make use of Botorch's GreedyVarianceReduction by setting the inducing point method to 'pivoted_chol', but we should be able to pass in allocator classes directly so that we can make use of GreedyImprovementReduction or any other methods that get implemented. The Sobol and kmeans++ heuristics should be implemented as InducingPointAllocator classes, and the select_inducing_points utility function should be deprecated.

The user should also be able to use the data itself as the inducing points.