Pass kwargs to post init in dataclasses

[zucchini-nlp]

As per title, needed for huggingface/transformers#41250. I verified that it is the easiest way to keep BC for remote code, because we can't force everyone to use dataclasses as configs

We don't want to set kwargs as attr directly, as some keys have to be popped (i.e. deprecated key) and some keys have to be transformed before setting (e.g. key attn_implememtation is set to private _attn_implememtation_internal)

While it can possibly be done in post init by manually deleting and post processing attr, imo it's cleaner to pass kwargs

# We could do smth like this right now, personally i find it ugly
def __post_init__(self):
    if hastattr(sefl, "deprecated_arg": del self.deprecated_arg
    if hastattr(sefl, "private_arg": self.move_to_private_property(private_arg)

# But after the suggested PR we don't need to do anything, and can keep existing code in transformers
def __post_init__(self, **kwargs):
        if (torch_dtype := kwargs.pop("torch_dtype", None)) is not None:
            self.dtype = self.dtype if self.dtype is not None else torch_dtype

[bot-ci-comment]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.56%. Comparing base (1daa48b) to head (95102bd).
⚠️ Report is 64 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #3771      +/-   ##
==========================================
+ Coverage   75.00%   76.56%   +1.55%     
==========================================
  Files         145      153       +8     
  Lines       13978    15086    +1108     
==========================================
+ Hits        10484    11550    +1066     
- Misses       3494     3536      +42     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Wauplin]

Thanks for the PR @zucchini-nlp! Makes perfect sense yes. I've updated the PR to add back the default behavior of setting kwargs as attributes but it can now be overwritten. I'll merge the PR now (failing CI is unrelated) and we'll ship it very soon :)