Introduced in #103488, they are used only in super_relate_tys.

I believe super_relate_tys should always relate types structurally. Any smarter inference rules should be handled by the caller instead, similar to how we do lazy normalization in the new solver.