When typechecking a type like
   Maybe (a :: <kind-sig>)
with a kind signature, we were using tc_lhs_kind to
typecheck the signature.  But that's utterly wrong; we
need the signature to be fully solid (non unresolved
equalities) before using it.  In the case of Trac #14904
we went on to instantiate the kind signature, when it
still had embedded unsolved constraints.  This tripped
the level-checking assertion when unifying a variable.

The fix looks pretty easy to me: just call tcLHsKind
instead.  I had to add KindSigCtxt to