Note [Prioritise class equalities]

We prioritise equalities in the solver (see selectWorkItem). But class
constraints like (a ~ b) and (a ~~ b) are actually equalities too;
see Note [The equality types story] in GHC.Builtin.Types.Prim.

Failing to prioritise these is inefficient (more kick-outs etc).
But, worse, it can prevent us spotting a "recursive knot" among
Wanted constraints.  See comment:10 of #12734 for a worked-out
example.

So we arrange to put these particular class constraints in the wl_eqs.

  NB: since we do not currently apply the substitution to the
  inert_solved_dicts, the knot-tying still seems a bit fragile.
  But this makes it better.