When we have an equality constraint where the LHS and RHS
have ill-matched kinds, it get turned into a CIrredEvCan
because a CTyEqCan/CFunEqCan are guaranteed kind-compatible.

But that in turn led to a bug because in the constraint
    c  =  (a:k1) ~ (b:k2)
the kind variables k1 and k2 don't show up in tyVarsOfType c.
Why not?  Because it looks like
    (~) k1 (a:k1) (b:k2)
Maybe (~) should have two kind arguments?  That seemed
like too big a change for not (we wait for NoKinds), so
this patch fixes the bug for now.