More flexible benchmark grouping

[ezyang]

Two use cases:

• We would like to add benchmarks over time / benchmarks get implicitly added/removed when they have correctness bugs fixed/introduced. Changing the set of running benchmarks affects the overall statistics over all benchmarks, which means we may incorrectly show an increase/decrease in stats, even though nothing materially changed (we just changed the measurement.)
• We have a lot of models, but sometimes there is a set of models for which we want to particularly focus on. For example, with dynamic shapes, there are a few models which we know will benefit from dynamic shapes in particular, because they have internal dynamism. We'd like to track against these models specifically.

Right now, benchmark grouping is hard-coded for the three benchmark suites. It would be nice if there were a light way to introduce new benchmark groupings on a model-by-model basis and see how they evolve over time. Some UI design necessary for managing these groupings.

cc @huydhn

[huydhn]

As this is not a simple fix IMO, I have added it into our sprint so that I can be planned accordingly

[ezyang]

a crappy mockup (like maybe just some manually drawn screens in excalidraw) I think would help us iterate on the UI design quickly, without having to build everything first