Note [fillInferResult]

When inferring, we use fillInferResult to "fill in" the hole in InferResult
   data InferResult = IR { ir_uniq :: Unique
                         , ir_lvl  :: TcLevel
                         , ir_ref  :: IORef (Maybe TcType) }

There are two things to worry about:

1. What if it is under a GADT or existential pattern match?
   - GADTs: a unification variable (and Infer's hole is similar) is untouchable
   - Existentials: be careful about skolem-escape

2. What if it is filled in more than once?  E.g. multiple branches of a case
     case e of
        T1 -> e1
        T2 -> e2

Our typing rules are:

* The RHS of a existential or GADT alternative must always be a
  monotype, regardless of the number of alternatives.

* Multiple non-existential/GADT branches can have (the same)
  higher rank type (#18412).  E.g. this is OK:
      case e of
        True  -> hr
        False -> hr
  where hr:: (forall a. a->a) -> Int
  c.f. Section 7.1 of "Practical type inference for arbitrary-rank types"
       We use choice (2) in that Section.
       (GHC 8.10 and earlier used choice (1).)

  But note that
      case e of
        True  -> hr
        False -> \x -> hr x
  will fail, because we still /infer/ both branches, so the \x will get
  a (monotype) unification variable, which will fail to unify with
  (forall a. a->a)

For (1) we can detect the GADT/existential situation by seeing that
the current TcLevel is greater than that stored in ir_lvl of the Infer
ExpType.  We bump the level whenever we go past a GADT/existential match.

Then, before filling the hole use promoteTcType to promote the type
to the outer ir_lvl.  promoteTcType does this
  - create a fresh unification variable alpha at level ir_lvl
  - emits an equality alpha[ir_lvl] ~ ty
  - fills the hole with alpha
That forces the type to be a monotype (since unification variables can
only unify with monotypes); and catches skolem-escapes because the
alpha is untouchable until the equality floats out.

For (2), we simply look to see if the hole is filled already.
  - if not, we promote (as above) and fill the hole
  - if it is filled, we simply unify with the type that is
    already there

(FIR1) There is one wrinkle.  Suppose we have
             case e of
                T1 -> e1 :: (forall a. a->a) -> Int
                G2 -> e2
    where T1 is not GADT or existential, but G2 is a GADT.  Then suppose the
    T1 alternative fills the hole with (forall a. a->a) -> Int, which is fine.
    But now the G2 alternative must not *just* unify with that else we'd risk
    allowing through (e2 :: (forall a. a->a) -> Int).  If we'd checked G2 first
    we'd have filled the hole with a unification variable, which enforces a
    monotype.

    So if we check G2 second, we still want to emit a constraint that restricts
    the RHS to be a monotype. This is done by ensureMonoType, and it works
    by simply generating a constraint (alpha ~ ty), where alpha is a fresh
unification variable.  We discard the evidence.