[23] VerifyError during conversion between T extends Short and short

[stephan-herrmann]

when checking primitive conversions involving a type variable, always use the superclass to find a boxing type, knowing that all boxing types are classes, so other bounds are irrelevant.

fixes #2937

[stephan-herrmann]

We have been using erasure() already in a few placed to find a boxing type hidden in a type variable, but we missed (at least) one relevant location.

OTOH, I don't think we have a card blanche to use erasure() for other situations, so I tried to establish the following principles:

• consistently use superclass() for the substitution rather than erasure() with the following justification:
  • we only look for boxing types among supers
  • since boxing types are classes, they must be found in superclass
• only perform such substitution if both are true:
  • the original type is a type variable (including capture)
  • the superclass is a boxing type indeed

The first item should not be a semantic change, just document that and why looking in this one location is sufficient.

I left one questionable use of erasure() in the context of NARROWING_AND_UNBOXING_CONVERSION which has a TODO anyway.