Instead of creating a new meta-tyvar and then unifying it with
a known kind in a KindedTyVar in a LHsQTyVars, just use the known kind.

Sadly, this still doesn't make #11719 usable, because while you can
now define a higher-kinded type family, you can't write any equations
for it, which is a larger problem.

test cases: dependent/should_compile/T{11635,11719}