Our EvaluationContext cache invalidation computation was consuming the
method token and version, but not the identity of the declaring assembly.
This lead to a strange bug where we would reuse the cache across assembly
boundaries if consecutive breakpoint happened to have the same method
token (each in its own assembly).

I believe this bug existed before I revised the cache invalidation
computation and we were just getting lucky - we used to check the spans
of all containing scopes, which are unlikely to match exactly across
assemblies.