This is allegedly causing large performance problems, see 13821

typeshed/8231 had zero hits on mypy_primer, so it's not the worst thing
to undo. Patching this in typeshed also feels weird, since there's a
more general soundness issue. If a typevar has a bound or constraint, we
might not want to solve it to a Literal.

If we can confirm the performance regression or fix the unsoundness
within mypy, I might pursue upstreaming this in typeshed.

(Reminder: add this to the sync_typeshed script once merged)