RequirementMachine: Add more limits to catch runaway computation, and fix a bug

[slavapestov]

The fuzzer found a couple of issues where we were missing termination limits to prevent runaway computation.

Also, we incorrectly accepted this protocol below, and proceeded to prove nonsense about it, because of a book-keeping bug in Knuth-Bendix completion. If two rules overlap at more than one position, we would only resolve the critical pair at the first position. Oops!

In a hypothetical world where this protocol is accepted, all of the same() calls should also be accepted because each equivalence holds under the protocol's same-type requirements. (You can work this out by hand even). Today, we diagnose that the last three calls to same() have mismatched argument types.

The correct behavior is to reject M3 because it doesn't actually have a finite complete presentation over any alphabet:

protocol M3 {
  associatedtype A: M3
  associatedtype B: M3
    where A.A.A == A, A.B.B.A == B.B  
}

func same<T>(_: T.Type, _: T.Type) {}

func f<T: M3>(_: T.Type) {
  same(T.A.A.A.self, T.A.self)
  same(T.A.B.B.A.self, T.B.B.self)
  same(T.B.B.A.self, T.A.B.B.self)
  same(T.A.A.B.B.self, T.B.B.self)
  same(T.A.A.B.A.B.B.self, T.B.A.B.B.self)
  same(T.A.A.B.A.B.A.B.B.self, T.B.A.B.A.B.B.self)
  same(T.A.A.B.A.B.A.B.A.B.B.self, T.B.A.B.A.B.A.B.B.self)  
}

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

@swift-ci Please test source compatibility

[slavapestov]

@swift-ci Please smoke test macOS