gh-105936: Properly update closure cells for __setattr__ and __detattr__ in frozen dataclasses with slots

[Prometheus3375]

Specifying frozen=True for a dataclass adds methods __setattr__ and __detattr__ to it. These methods reference the class via closure. If slots=True is also present, then the class is recreated with slots with all methods being copied. Notably, every method referencing the old class via closure in variable __class__ is properly updated, but __setattr__ and __detattr__ reference the old class via cls; hence, they are still referencing the old class instead of the new one. This PR fixes this issue.

• Issue: dataclasses: frozen dataclasses with slots' __setattr__ and __delattr__ does not work. #105936

Additional details:

• 3.14 regression: slot dataclasses classes leak original class #135228 (comment)
• 3.14 regression: slot dataclasses classes leak original class #135228 (comment)

The issue is present in Python as early as in 3.10. This PR must be backported to 3.14 and possibly 3.13.

Originally this PR was referencing gh-135228, but soon I realized that the issue is not connected to regression in 3.14. There is an older PR #105937 that fixes the issue, but it is heavily outdated.

[python-cla-bot]

All commit authors signed the Contributor License Agreement.

[JelleZijlstra]

Thanks! For the test you can do something like that in the original fix, https://github.com/python/cpython/pull/137047/files#diff-212e368b34eb9b134f87e765787d6d26b747235a358e13da8922f83861c0d676 .

[Prometheus3375]

While this PR is not merged, should AKCS be updated?

[JelleZijlstra]

If I remember correctly the latest policy on Misc/ACKS is that people can add themselves if they want but it's not required. The file is a pain to keep fully correct and sort of redundant with git log, so there's been talk of getting rid of it but not sure anything concrete has changed yet.