Note [Unifying coercion-foralls]

Suppose we try to unify (forall cv. t1) ~ (forall cv. t2).
See Note [ForAllTy] in GHC.Core.TyCo.Rep.

The problem with coercion variables is that coercion abstraction is not erased:
the `kco` shouldn't propagate from outside the ForAllTy to inside. Instead, I think
the correct new `kco` for the recursive call is `mkNomReflCo liftedTypeKind` (but I'm
a little worried it might be Constraint sometimes).

This potential problem has been there a long time, and I'm going to let
sleeping dogs lie for now.