Reconsider supports of constraint attributes in bridges

See discussion in #1992, and these tests:
https://github.com/jump-dev/MathOptInterface.jl/blob/d9c998ac49a80ed8f740d5009133320e982af977/test/Bridges/bridge_optimizer.jl#L898-L933

The issue is that the bridge/model may support an attribute for a variable-in-set constraint if it was added via `add_constrained_variable(s)` but not via `add_constraint` (or vice versa). However, we have no way of telling which way the constraint was added based on the type alone.

We can either default to `&&`, in which case we may return a false negative (the model supports the attribute but `supports` returns `false`) or `||`, in which case we may return a false positive (the model supports the attribute for one of the ways, but we added it the other way). It's not obvious which is better, but our tests are passing with the current implementation, so we should wait for a solver to complain before reconsidering.