Fix free-threaded cleanup race in _replace_dunder_methods

[Furkan-rgb]

What does this PR do?

Fixes a race in lightning.fabric.utilities.data._replace_dunder_methods when the cleanup path runs concurrently under free-threaded Python.

In the finally block, two overlapping callers can observe __old__* in cls.__dict__, but one thread may restore and delete the attribute before the other finishes its own getattr/setattr/delattr sequence. Under Python 3.14t with PYTHON_GIL=0, that can raise AttributeError and interrupt the training loop.

This change makes the cleanup tolerant to that race by skipping AttributeError if another thread has already restored the original dunder method. It also adds a regression test that simulates a concurrent restore during __old__delattr__ cleanup.

Why is this needed?

This affects free-threaded Python runs where Lightning patches overlapping class hierarchies from multiple threads, for example threaded hyperparameter searches.

Does this PR introduce any breaking changes?

No.

How was this tested?

• python -m pytest tests/tests_fabric/utilities/test_data.py -k replace_dunder_methods -v
• python -m pytest tests/tests_pytorch/utilities/test_data.py::test_custom_torch_batch_sampler -v

---

📚 Documentation preview 📚: https://pytorch-lightning--21618.org.readthedocs.build/en/21618/

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 87%. Comparing base (6805188) to head (f88aca9).
⚠️ Report is 14 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master   #21618   +/-   ##
=======================================
- Coverage      87%      87%   -0%     
=======================================
  Files         270      270           
  Lines       23895    23899    +4     
=======================================
+ Hits        20675    20676    +1     
- Misses       3220     3223    +3