[Sema] Extend fix for reference storage types in protocol witnesses to include generics.

[gregomni]

Had wrongly disregarded generic conformers in my original fix for sr-6097. This commit extends that fix.

(I had only been thinking in terms of methods with generic parameters before (which obviously couldn't contain reference storage types) -- oops.)

[gregomni]

@swift-ci Please smoke test

[gregomni]

@slavapestov Should've asked for review on the last one, sorry... can you take a quick glance at this?

As far as the deduction in the tests -- the indication of regression would be errors during compilation, a la:

gen.swift:5:7: error: type 'G<T>' does not conform to protocol 'SomeProto'
class G<T> : SomeProto where T : AnyObject {
      ^
gen.swift:2:18: protocol requires nested type 'A'; do you want to add it?

If you did somehow get past Sema with a conformance where the associated type witness was still weak/unowned you'd run into asserts in Silgen in generating the witness table functions, because Silgen doesn't know how to do those wrong type transformations. (I went down this blind alley for a bit.)

[slavapestov]

This looks good, thanks!

Indeed we don't want associated type witnesses to have weak or unowned types. At the SILGen level, a storage requirement is witnessed by the getter/setter pair, and the type of the getter and setter does not include weak or unowned. Essentially those are not types at all, but a property of how the underlying storage is implemented.