It improves compile time in `--release` mode quite a bit, it doesn't
really slow things down and, conceptually, it seems closer to what we
want the physical architecture to look like (we don't want to
monomorphise EVERYTHING in a single leaf crate).
To test whether the receiver type matches for the impl, we unify the given self
type (in this case `HashSet<{unknown}>`) with the self type of the
impl (`HashSet<?0>`), but if the given self type contains Unknowns, they won't
be unified with the variables in those places. So we got a receiver type that
was different from the expected one, and concluded the impl doesn't match.
The fix is slightly hacky; if after the unification, our variables are still
there, we make them fall back to Unknown. This does make some sense though,
since we don't want to 'leak' the variables.
Fixes#3547.
Basically, we need to allow variables in the caller self type to unify with the
impl's declared self type. That requires some more contortions in the variable
handling. I'm looking forward to (hopefully) handling this in a cleaner way when
we switch to Chalk's types and unification code.
The stand-alone `unify` requires that the type doesn't contain any type
variables. So we can't share the code here for now (without more refactoring)...
