We were recovering too eagerly from errors in pattern-synonym
type inference, leading to a cascade of confusing follow up errors
(Trac #15685, #15692).

The underlying issue is that a pattern synonym should have a closed,
fixed type, with no unification variables in it.  But it wasn't!

Fixing this made me change the interface to simplifyInfer slightly.
Instead of /emitting/ a residual implication constraint, it
now /returns/ it, so that the caller can decide what to do.