Fix #9463: create varargs forwarder symbols in transform instead of transformDefDef

[TheElectronWill]

This allows the forwarder symbols to be visible during separate compilation.
See smarter's comment in #9463.

[smarter]

Looks like this is missing a pos test with the code from #9463 ?

[TheElectronWill]

Oopsie 👀 Here are the tests.
I've moved (or deleted, when the use case was already covered by another test) some disabled varargs tests, they all pass now.

[TheElectronWill]

I don't understand why tests/pos/annot-bootstrap (which I didn't touch) fails now, especially as it's not consistent: sometimes it fails in test, sometimes in test_bootstraped, sometimes both, and I can't reproduce it locally 😕
When looking for the overriden symbols in overridesJava in transform (here), it ends up calling complete on a NoLoader, which fails.

[smarter]

Hmm, not sure what's going on. One thing I see is that overridesJava(sym) is now going to be called more often since the ref1 ne ref check was moved below it, so I guess this is causing some sort of cycle, but only in very specific conditions? From the stacktrace I see that the call to complete the NoLoader symbol comes from isAbsent which is itself called from TreeChecker#transformSym, that call could be changed to isAbsent(canForce = false) to avoid that completion (this won't break anything since TreeChecker is just here to run extra sanity checks anyway).

[TheElectronWill]

🤔 Hmmm canForce = false didn't fix the problem, it created more errors. But I have another idea.

[TheElectronWill]

It works! 🎉

[smarter]

Originally the ref1 ne ref was here which I think is equivalent to what the code is doing now?

[TheElectronWill]

Not quite, because now transformVarArgs is called for all methods. If I move ref1 ne ref back to the case guard, some situations are not covered and e.g. tests/neg/varargs-annot fails.

[smarter]

This isn't new to this PR, but I've noticed something weird with this phase: it extends InfoTransformer, which are supposed to only change the info of denotation and leave everything else alone, but in this phase we actually override transform to do more stuff on top of that, so it's a bit misleading. I think it'd be good to rewrite as a regular DenotTransformer, especially now that transform has gotten more complicated.