Sema: Fix request cycle involving protocol typealiases in preparation for RequirementMachine support

[slavapestov]

The RequirementMachine eagerly resolves protocol typealiases while building a rewrite system. This was causing a request cycle because structural type resolution would request the generic signature of a type unnecessarily, and that generic signature could then reference the protocol itself.

Here is an example (my next PR actually adds an explicit test case, but there are a bunch of examples in the test suite already that boil down to this and fail with -requirement-machine-protocol-signatures=on):

protocol P where A == B {
  associatedtype A
  typealias B = Foo<Self>
}

struct Foo<T : P> {}

While building the requirement signature of P we resolve the structural type of B, which would previously trigger the computation of the generic signature of Foo, which would ask for the requirement signature of P. However structural type resolution doesn't check generic requirements on the bound generic type, so it need the generic signature except for in a couple of specific cases (which will still cause a cycle with an example like the above, but that's okay).

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please test source compatibility

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please smoke test macOS