Fix upwards closure for relations (#7515)

Fix upwards closure for relations (#7515)

This PR fixes an issue where we did compute upwards closure (for join, product, etc.) on equivalence classes whose membership cache had already been initialized (perhaps recursively from the computation of upwards/downwards closures on other terms).

It also generalizes the fix from #7511. Instead of doing explicit splitting, we mark shared terms and let theory combination (when necessary) do splitting. This is required to fix a more general version of the benchmark from the previous PR, where instead now a term c+1 is used in a position that is relevant to a join.

It disables a regression that times out after these fixes, and does further cleanup.