gh-133956 fix bug where dataclass wouldn't detect ClassVar fields if ClassVar was re-exported from a module other than typing

[taminomara]

This PR continues work by @dzherb from #134073.

I've simplified the code and removed strict checks from _is_type. Supporting arbitrary levels of module nesting turned out to be beneficial for overall code complexity, so I've included it to this PR; I can send it as a separate PR with a separate issue, though.

Motivation for current implementation

I think we shouldn't assume anything about user-provided type annotations. All we know from annotation ty.ClassVar is that there is an object ty, and we're getting its attribute ClassVar. We shouldn't assume that ty is a module, nor that it doesn't implement some custom __getattr__ logic; instead, we should just do what the annotation tells us to do.

• Issue: dataclasses: Synthetic __init__ method on dataclass has broken ClassVar annotation under some conditions. #133956

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[emcd]

If someone is interested in confirming this works, writing tests, and putting up a PR, I'd be happy to review it.

@JelleZijlstra : In May, you said that you would be happy to review a PR which contains the fix and has tests. I believe that this PR meets those criteria. Would you review it?

[taminomara]

@JelleZijlstra are there any updates on this?

[emcd]

@taminomara : You might also want to try pinging people in the issue thread too. They might ignore PRs if they do not have enough context.

Thank you for making this patch. I briefly glanced at it when you first made it a couple months ago and it looked like what I had in mind. Happy holidays!

[taminomara]

Rebased onto current main, no functional changes. Pre-rebase changeset was 8a855b4, saved in gh-133956-broken-classvar-pre-rebase.