The following no longer crashes:

    class C { type F[X <: F[_, _], Y] }

But it is still flagged as a cyclic reference error, whereas nsc accepts it. I believe it would be
tricky/risky to change dotc's algorithms to accept it as well, and I am not sure it's
necessary. Generally, we want to get away from F-bounds, so just accepting the most
common use case (F-bounds in method type parameters) is hopefully OK.