Note [Add signature contexts as wanteds]

Consider this (#11016):
  f2 :: (?x :: Int) => _
  f2 = ?x

or this
  class C a b | a -> b
  g :: C p q => p -> q
  f3 :: C Int b => _
  f3 = g (3::Int)

We'll use plan InferGen because there are holes in the type.  But:
 * For f2 we want to have the (?x :: Int) constraint floating around
   so that the functional dependencies kick in.  Otherwise the
   occurrence of ?x on the RHS produces constraint (?x :: alpha), and
   we won't unify alpha:=Int.

 * For f3 want the (C Int b) constraint from the partial signature
   to meet the (C Int beta) constraint we get from the call to g; again,
   fundeps

Solution: in simplifyInfer, we add the constraints from the signature
as extra Wanteds.

Why Wanteds?  Wouldn't it be neater to treat them as Givens?  Alas
that would mess up (GivenInv) in Note [TcLevel invariants].  Consider
    f :: (Eq a, _) => blah1
    f = ....g...
    g :: (Eq b, _) => blah2
    g = ...f...

Then we have two psig_theta constraints (Eq a[tv], Eq b[tv]), both with
TyVarTvs inside.  Ultimately a[tv] := b[tv], but only when we've solved
all those constraints.  And both have level 1, so we can't put them as
Givens when solving at level 1.

Best to treat them as Wanteds.

But see also #20076, which would be solved if they were Givens.


************************************************************************
*                                                                      *
                Quantification
*                                                                      *
************************************************************************