This check is somewhat subtle. See Note [Checking GADT return types]
in TcTyClsDecls. The new plan is to check *before* desugaring the type
from HsType to Type. This avoids problems with the pattern-match
in rejigConRes.

As a nice side benefit to this, I discovered that Template Haskell
splices were a little conservative in their treatment of valid data
constructors. (For example, a kind signature in the return type caused
failure.) Now, the TH code uses exactly the same function as the
"real" code, which is nice. See hsTyGetAppHead_maybe in HsTypes.