This change tidies up and simplifies (a bit) the knot-tying
when kind-checking groups of type and class declarations.

The trouble (shown by Trac #11356) was that we wanted an error message
(a kind-mismatch) that involved a type mentioned a (AThing k), which
blew up.

Since we now seem to have TcTyCons, I decided to use them here.
It's still not great, but it's easier to understand and more robust.