Fix regression around type skolems and if exprs.

[retronym]

If we start with:

async({
  val res = await[S[_$1 with String]](s);
  if (true)
    await[Int](0)
  res
})

Typechecking the application (before macro expansion) yields
(where the trees are printed in the form expr{tpe}):

async[S[_$1#5738 with String#137]]({
  val res: S[_$1#5490 with String] forSome { type _$1#5490 } =
    await[S[_$1#5487 with String]](
      s{S[_$1#5487 with String]}
    ){S[_$1#5487 with String]};
  if (true)
    await(0)
  else
    ()
  res{S[_$1#5738 with String]}
}{S[_$1#5738 with String]}){S[_$1#5738 with String]}

Note that the type of the second last line contains a skolemized
symbol _$1#5738 of the existential _$1#5490. This is created
by this case in Typer#adapt:

case et @ ExistentialType(_, _) if ((mode & (EXPRmode | LHSmode)) == EXPRmode) =>
  adapt(tree setType et.skolemizeExistential(context.owner, tree), mode, pt, original)

Our ANF rewrites part of this code to:

<synthetic> val await$1: S[_$1#5487 with String] = await[S[_$1#5487 with String]](awaitable$1);
val res: S[_$1#5490 with String] forSome { type _$1 } = await$1;

And later, the state machine transformation splits the last line into
a blank field and an assignment. Typechecking the Assign node
led to the an type error.

This commit manually attributes the types to the Assign node so
as to avoid these problem.

It also reigns in an overeager rewriting of If nodes in the
ANF transform, which was due to a bug in the label detection
logic introduced in 4fc5463.

Thanks to @gnovark for yet another devilish test case and
analysis of the problem with label detection.

[retronym]

Review by @lrytz

Supercedes #77, Fixes #79

[retronym]

It looks like there could be a bug (or limitation) of existential subtyping in Scala that makes it hard for us to take a principled approach here.

I've prototyped a fix here: https://github.com/retronym/scala/tree/topic/existential-in-compound. See the test case in that branch for more details.

Here's a sketch of the way that async ought to type the synthetic vals introduced by the ANF transform: retronym@eec1cab

We can still pursue this approach and insert casts where we determine that the repacked existential type on the LHS of an assignment does't accomodate the skolemized existential in the RHS.