Try to instantiate type variables in tryInsertImplicitOnQualifier

[odersky]

Fixes #13842

What happened in #13842 was that we were not hitting this case in
tryWiden, which is a method in ApproximatingTypeMap that tries
to dealias or widen before propagating a Range outwards:

          case info: SingletonType =>
            // if H#x: y.type, then for any x in L..H, x.type =:= y.type,
            // hence we can replace with y.type under all variances
            reapply(info)

The actual info here was a TypeVar with a singleton type as lower bound. Instantiating
the TypeVar produces a SingletonType and the case applies.

We cannot really instantiate TypeVars here since this is very low-level code. For instance
we might be in a state where the instantiation is provisional and the TyperState is thrown
away. But AsSeenFrom results are cached so we'd have the wrong type in the caches.

What we do instead is add an additional case in tryInsertImplicitOnQualifier which is the
last resort when an application fails. Here, if we can instantiate some type variables in the
qualifier, we try again.

[smarter]

But AsSeenFrom results are cached so we'd have the wrong type in the caches.

But we make sure to not cache asSeenFrom when the prefix isProvisional: https://github.com/lampepfl/dotty/blob/4ad416fc3153c953835067e25c6a12309f714256/compiler/src/dotty/tools/dotc/core/Denotations.scala#L137 and an instantiation of a type variable won't set isProvisional to false unless the instantiation is permanent, so I think this would be fine?

[odersky]

I still think TypeMap is too low-level to be in the business of instantiating type variables. TypeMaps should not have observable side effects.