gh-135228: 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.

More details in these comments: 1, 2.

Any help with tests for the issue would be appreciated. I guess just checking that __setattr__ and __delattr__ reference the new class would be sufficient.

• Issue: 3.14 regression: slot dataclasses classes leak original class #135228

[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]

Added 3 tests:

1. That the original class is gc-ed if it is a frozen slotted dataclass.
2. That methods added by frozen=True reference the new class via __class__.
3. That methods added by frozen=True do not reference the old class.

I will add news a bit later.

[Prometheus3375]

By the way, 3.12 also suffers from the same issue. I do not have 3.13, but I guess there it is also present. This PR should be backported to 3.13.