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 implication and then gets floated out from the implication,
the Derived is present both in and out of the implication. This
causes a duplicate error message. (Test case:
typecheck/should_fail/T7368) Solution: track the provenance of
Derived constraints and refuse to float out a constraint that has
an insoluble Derived.

Lastly, this labels one test (dependent/should_fail/RAE_T32a)
as expect_broken, because the problem is really #12919. The
different handling of constraints in this patch exposes the error.

This fixes bugs #11198, #12373, #13530, and #13610.

test cases:
typecheck/should_fail/{T8262,T8603,tcail122,T12373,T13530,T13610}