Note [Decomposing type family applications]

Supose we have
   [G/W]  (F ty1) ~r  (F ty2)
This is handled by the TyFamLHS/TyFamLHS case of canEqCanLHS2.

We never decompose to
   [G/W]  ty1 ~r' ty2

Instead

* For Givens we do nothing. Injective type families have no corresponding
  evidence of their injectivity, so we cannot decompose an
  injective-type-family Given.

* For Wanteds, for the Nominal role only, we emit new Wanteds rather like
  functional dependencies, for each injective argument position.

  E.g type family F a b   -- injective in first arg, but not second
      [W] (F s1 t1) ~N (F s2 t2)
  Emit new Wanteds
      [W] s1 ~N s2
  But retain the existing, unsolved constraint.