Relax Note [Phantom type variables in kinds] to allow for phantom levity vars

While implementing the unlifted datatypes proposal (!2218 (closed)), I realised that an edge case violates the kinding rules (Note [Phantom type variables in kinds], in particular). Specifically

type T :: forall (l :: Levity). TYPE (BoxedRep l)
data T = MkT

Note that

1. T has a nullary data con MkT
2. T is levity-polymorphic

That means the type of MkT, forall {l::Levity}. T @l, is ill-kinded! We have T @l :: TYPE (BoxedRep l) and the forall is supposed to bind the l. But the kind rule for forall a. TYPE rep demands that a is not free in rep:

          ty : TYPE rep
          `a` is not free in rep
(FORALL1) -----------------------
          forall a. ty : TYPE rep

I argue that there is no harm in allowing TYPE (BoxedRep l) here, because l does not affect the representation of T. I'm unaware of any other implications this might have.

We have to be a bit careful with evaluation semantics, but we have better means (#15532 (comment 239845), #17521) to decide whether levity polymorphism is OK at the term level. In any instance, MkT is a data constructor and thus is a value under both call-by-need and call-by-value semantics. I'm reasonably positive that allowing MkT to be well-kinded is OK. You can't do anything with it at the term level unless you apply it to a Levity anyway.