AST/Sema: Fix a couple of minor issues with tuple conformances and add a new test case
The progress on variadic generics means we can now implement useful
witnesses in a tuple conformance. The feature remains very incomplete
though, today we crash in SILGen.