This concretely improves type inference of some cases (see included
test). I assume the compiler struggles to reason about multiple layers
of generic type parameters (even with associated-type equalities) but
*can* understand pure associated types, since they are always directly
computable from the input types.