Note [Improvement orientation]

See also Note [Fundeps with instances, and equality orientation], which describes
the Exact Same Problem, with the same solution, but for functional dependencies.

A very delicate point is the orientation of equalities
arising from injectivity improvement (#12522).  Suppose we have
  type family F x = t | t -> x
  type instance F (a, Int) = (Int, G a)
where G is injective; and wanted constraints

  [W] F (alpha, beta) ~ (Int, <some type>)

The injectivity will give rise to constraints

  [W] gamma1 ~ alpha
  [W] Int ~ beta

The fresh unification variable gamma1 comes from the fact that we
can only do "partial improvement" here; see Section 5.2 of
"Injective type families for Haskell" (HS'15).

Now, it's very important to orient the equations this way round,
so that the fresh unification variable will be eliminated in
favour of alpha.  If we instead had
   [W] alpha ~ gamma1
then we would unify alpha := gamma1; and kick out the wanted
constraint.  But when we substitute it back in, it'd look like
   [W] F (gamma1, beta) ~ fuv
and exactly the same thing would happen again!  Infinite loop.

This all seems fragile, and it might seem more robust to avoid
introducing gamma1 in the first place, in the case where the
actual argument (alpha, beta) partly matches the improvement
template.  But that's a bit tricky, esp when we remember that the
kinds much match too; so it's easier to let the normal machinery
handle it.  Instead we are careful to orient the new
equality with the template on the left.  Delicate, but it works.