Experiment should merge redundant feature group subsets

If the user creates a bunch of feature group subsets, there can be some redundancy. We should do some mixture of:

- blowing up in the experiment validator if they do configuration that makes no sense (like mixing all-combinations with anything else)
- merging redundant subsets if they are arrived at through sane config (e.g. leave-one-in and leave-one-out overlap)

Start by locating the experiment validator and the code that constructs feature-group subsets. Determine which configurations are invalid, including mixing all-combinations with other modes, and how leave-one-in and leave-one-out overlap. Done means invalid configurations fail validation and valid redundant subsets are merged, with tests covering both cases.

Written by the indexing model from the issue text.