Restrict Eta-reduction to Java-defined classes

[odersky]

Collapse [X] =>> C[X] only if C is a Java-defined class. Since
Java does not have hk-types it must be a raw type in this case.

[smarter]

We try to always cook raw types so in what case is this actually needed ?

[odersky]

@smarter Failing tests if the case EtaExpansion clauses are completely removed:

    tests/pos/t3349 failed
    tests/pos/t3429 failed
    tests/pos-java-interop/t1196 failed

[smarter]

OK, so it seems the issues all come from joint-compilation with .java sources where we don't actually cook types.

[odersky]

OK, we cook types coming from Java sources now as well, which means we can simplify TypeComparer to no longer test for eta expansions.

[odersky]

Backing out. It seems we cannot cook things everywhere without running into cyclic references.