Note [Deciding quantification]

If the monomorphism restriction does not apply, then we quantify as follows:

* Step 1: decidePromotedTyVars.
  Take the global tyvars, and "grow" them using functional dependencies
     E.g.  if x:alpha is in the environment, and alpha ~ [beta] (which can
          happen because alpha is untouchable here) then do not quantify over
          beta, because alpha fixes beta, and beta is effectively free in
          the environment too; this logic extends to general fundeps, not
          just equalities

  We also account for the monomorphism restriction; if it applies,
  add the free vars of all the constraints.

  Result is mono_tvs; we will promote all of these to the outer levek,
  and certainly not quantify over them.

* Step 2: defaultTyVarsAndSimplify.
  Default any non-promoted tyvars (i.e ones that are definitely
  not going to become further constrained), and re-simplify the
  candidate constraints.

  Motivation for re-simplification (#7857): imagine we have a
  constraint (C (a->b)), where 'a :: TYPE l1' and 'b :: TYPE l2' are
  not free in the envt, and instance forall (a::*) (b::*). (C a) => C
  (a -> b) The instance doesn't match while l1,l2 are polymorphic, but
  it will match when we default them to LiftedRep.

  This is all very tiresome.

  This step also promotes the mono_tvs from Step 1. See
  Note [Promote monomorphic tyvars]. In fact, the *only*
  use of the mono_tvs from Step 1 is to promote them here.
  This promotion effectively stops us from quantifying over them
  later, in Step 3. Because the actual variables to quantify
  over are determined in Step 3 (not in Step 1), it is OK for
  the mono_tvs to be missing some variables free in the
  environment. This is why removing the psig_qtvs is OK in
  decidePromotedTyVars. Test case for this scenario: T14479.

* Step 3: decideQuantifiedTyVars.
  Decide which variables to quantify over, as follows:

  - Take the free vars of the partial-type-signature types and constraints,
    and the tau-type (zonked_tau_tvs), and then "grow"
    them using all the constraints.  These are grown_tcvs.
    See Note [growThetaTyVars vs closeWrtFunDeps].

  - Use quantifyTyVars to quantify over the free variables of all the types
    involved, but only those in the grown_tcvs.

  Result is qtvs.

* Step 4: Filter the constraints using pickQuantifiablePreds and the
  qtvs. We have to zonk the constraints first, so they "see" the
  freshly created skolems.