Sema: Properly handle inference for abstract fixed type witnesses

[AnthonyLatsis]

This fixes inference for some fixed type witness chains - SR-10831, for example:

protocol P {
  associatedtype A
  associatedtype B
}
protocol Q: P where A == G<B> {}
protocol R: Q where B == Int {}

struct Foo: R {} // OK

And adds inference of fixed type witnesses containing Self - SR-11671:

protocol P {
  associatedtype A
}
protocol Q: P where A == Self {}

struct Foo: Q {} // OK, Self == Foo

@slavapestov

[AnthonyLatsis]

@slavapestov bumping this up in case you have some bandwidth

[AnthonyLatsis]

After almost a year I can tell the approach here is flawed and quite narrow. Closing to revisit this once we settle on the implementation details of a new conformance checker.