Handle more corner cases in etaReduce

[odersky]

Generalize eta reduce so that it also handles eta expansions that are subject to
adaptation or specialization.

Fixes #14623

[odersky]

@smarter One tricky aspect is that eta reduction has to undo closure adaptation byErasure#adaptClosure. When looking at this code, it seemed that there should be a simpler way. We have a closure and create another closure that calls the original closure and adapts its parameters or result. Why not modify the existing closure instead?

[odersky]

Btw doing eta reduction before erasure would not be a solution here since the eta expansion in #14623 does change the result type (from normal function to context function). It's only after erasure that we can assume the two are the same.

[smarter]

We have a closure and create another closure that calls the original closure and adapts its parameters or result. Why not modify the existing closure instead?

Hmm, this might be possible by special-casing the erasure of the parameter symbols of closure defs in TypeErasure#eraseInfo, though it's not clear to me if that would be a net win in terms of complexity or performance, in particular it seems like we could end up with many more boxing/unboxing steps.