Improve error messages around kind mismatches.
Previously, when canonicalizing (or unifying, in uType) a heterogeneous equality, we emitted a kind equality and used the resulting coercion to cast one side of the heterogeneous equality. While sound, this led to terrible error messages. (See the bugs listed below.) The problem is that using the coercion built from the emitted kind equality is a bit like a wanted rewriting a wanted. The solution is to keep heterogeneous equalities as irreducible. See Note [Equalities with incompatible kinds] in TcCanonical. This commit also removes a highly suspicious switch to FM_SubstOnly when flattening in the kinds of a type variable. I have no idea why this was there, other than as a holdover from pre-TypeInType. I've not left a Note because there is simply no reason I can conceive of that the FM_SubstOnly should be there. One challenge with this patch is that the emitted derived equalities might get emitted several times: when a heterogeneous equality is in an impl...