Fix union simplification performance regression

[JukkaL]

#11962 can generate large unions with many Instance types with
last_known_value set. This caused our union simplification algorithm
to be extremely slow, as it hit an O(n**2) code path.

We already had a fast code path for unions of regular literal types. This
generalizes it for unions containing Instance types with last known
values (which behave similarly to literals in a literal type context).

Also fix a union simplification bug that I encountered while writing tests
for this change.

Work on #12408.

[github-actions]

According to mypy_primer, this change has no effect on the checked open source code. 🤖🎉