Note [Error on unconstrained meta-variables]

Consider

* type C :: Type -> Type -> Constraint
  class (forall a. a b ~ a c) => C b c

* type T = forall a. Proxy a

* data (forall a. a b ~ a c) => T b c

* type instance F Int = Proxy Any
  where Any :: forall k. k

In the first three cases we will infer a :: Type -> kappa, but then
we get no further information on kappa. In the last, we will get
  Proxy kappa Any
but again will get no further info on kappa.

What do do?
 A. We could choose kappa := Type. But this only works when the kind of kappa
    is Type (true in this example, but not always).
 B. We could default to Any.
 C. We could quantify.
 D. We could error.

We choose (D), as described in #17567, and implement this choice in
doNotQuantifyTyVars.  Discussion of alternatives A-C is below.

NB: this is all rather similar to, but sadly not the same as
    Note [Naughty quantification candidates]

To do this, we must take an extra step before doing the final zonk to create
e.g. a TyCon. (There is no problem in the final term-level zonk. See the
section on alternative (B) below.) This extra step is needed only for
constructs that do not quantify their free meta-variables, such as a class
constraint or right-hand side of a type synonym.

Specifically: before the final zonk, every construct must either call
quantifyTyVars or doNotQuantifyTyVars. The latter issues an error
if it is passed any free variables. (Exception: we still default
RuntimeRep and Multiplicity variables.)

Because no meta-variables remain after quantifying or erroring, we perform
the zonk with NoFlexi, which panics upon seeing a meta-variable.

Alternatives A-C, not implemented:

A. As stated above, this works only sometimes. We might have a free
   meta-variable of kind Nat, for example.

B. This is what we used to do, but it caused Any to appear in error
   messages sometimes. See #17567 for several examples. Defaulting to
   Any during the final, whole-program zonk is OK, though, because
   we are completely done type-checking at that point. No chance to
   leak into an error message.

C. Examine the class declaration at the top of this Note again.
   Where should we quantify? We might imagine quantifying and
   putting the kind variable in the forall of the quantified constraint.
   But what if there are nested foralls? Which one should get the
   variable? Other constructs have other problems. (For example,
   the right-hand side of a type family instance equation may not
   be a poly-type.)

   More broadly, the GHC AST defines a set of places where it performs
   implicit lexical generalization. For example, in a type
   signature

     f :: Proxy a -> Bool

   the otherwise-unbound a is lexically quantified, giving us

     f :: forall a. Proxy a -> Bool

   The places that allow lexical quantification are marked in the AST with
   HsImplicitBndrs. HsImplicitBndrs offers a binding site for otherwise-unbound
   variables.

   Later, during type-checking, we discover that a's kind is unconstrained.
   We thus quantify *again*, to

     f :: forall {k} (a :: k). Proxy @k a -> Bool

   It is this second quantification that this Note is really about --
   let's call it *inferred quantification*.
   So there are two sorts of implicit quantification in types:
     1. Lexical quantification: signalled by HsImplicitBndrs, occurs over
        variables mentioned by the user but with no explicit binding site,
        suppressed by a user-written forall (by the forall-or-nothing rule,
        in Note [forall-or-nothing rule] in GHC.Hs.Type).
     2. Inferred quantification: no signal in HsSyn, occurs over unconstrained
        variables invented by the type-checker, possible only with -XPolyKinds,
        unaffected by forall-or-nothing rule
   These two quantifications are performed in different compiler phases, and are
   essentially unrelated. However, it is convenient
   for programmers to remember only one set of implicit quantification
   sites. So, we choose to use the same places (those with HsImplicitBndrs)
   for lexical quantification as for inferred quantification of unconstrained
   meta-variables. Accordingly, there is no quantification in a class
   constraint, or the other constructs that call doNotQuantifyTyVars.