(More) Declaration checker support for tuple conformances

[slavapestov]

This PR implements a new syntax that doesn't require -parse-stdlib, suggested by @hborla. The nice thing is that we want it to just work anyway, even if/when a new syntax is added:

typealias MyTuple<each Horse> = (repeat each Horse)
extension MyTuple: P where repeat each Horse: P {...}

Or with requirement inference:

typealias MyTuple<each Horse: P> = (repeat each Horse)
extension MyTuple: P {...}

This PR also implements enforcement of various semantic restrictions:

• Extended type must be (repeat each T) where T is the type parameter pack of the alias.

• A tuple extension must define a conformance to exactly one protocol P, this conformance must be conditional, and there must be exactly one conditional requirement that is a pack conformance requirement of T to P. You cannot have extension Tuple: P where repeat each T: Q, because if T is replaced with a single scalar type X, we get X: P where X: Q, which is false in general unless P inherits from Q.

• An associated type A must be witnessed by (repeat (each T).A) in the conformance; that is, we require that (repeat each T).A == (repeat (each T).A). This restriction is required for a similar reason, for coherence with the one-element vanishing substitution.

The above is not enough to ensure coherence. We also need to statically and dynamically unwrap one-element tuple conformances, and prevent devirtualization of certain calls. This will happen in a subsequent PR.

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please test Windows