Fix ddp_notebook CUDA fork check to allow passive initialization

[arrdel]

What does this PR do?

Fixes #21389

This PR fixes the overly strict CUDA fork check in ddp_notebook strategy that was causing false positives in notebook environments like Kaggle.

Problem

The previous implementation used torch.cuda.is_initialized() which returns True even when CUDA is passively initialized (e.g., during library imports, device availability checks, or model loading). This caused the error:

RuntimeError: Lightning can't create new processes if CUDA is already initialized.

This happened even when users didn't explicitly call any CUDA functions, making it impossible to use ddp_notebook in many legitimate scenarios.

Solution

This fix uses PyTorch's internal torch.cuda._is_in_bad_fork() function, which more accurately detects when we're in an actual bad fork state.

The implementation includes a fallback to the old check for older PyTorch versions that don't have _is_in_bad_fork.

Testing

• [x] Code follows style guidelines
• [x] Changes preserve backward compatibility
• [x] Fallback exists for older PyTorch versions

---

📚 Documentation preview 📚: https://pytorch-lightning--21402.org.readthedocs.build/en/21402/

[codecov[bot]]

Codecov Report

:x: Patch coverage is 0% with 11 lines in your changes missing coverage. Please review. :white_check_mark: Project coverage is 79%. Comparing base (79ffe50) to head (f002d00). :warning: Report is 13 commits behind head on master. :white_check_mark: All tests successful. No failed tests found.

:exclamation: There is a different number of reports uploaded between BASE (79ffe50) and HEAD (f002d00). Click for more details.

HEAD has 3345 uploads less than BASE

Flag	BASE (79ffe50)	HEAD (f002d00)
cpu	777	30
lightning_fabric	195	0
pytest	390	0
python3.12	233	9
python3.12.7	232	9
lightning	388	15
python3.11	156	6
python3.10	78	3
python	78	3
pytorch2.1	78	6
pytest-full	387	30
pytorch_lightning	194	15
pytorch2.6	39	3
pytorch2.4.1	38	3
pytorch2.3	39	3
pytorch2.2.2	39	3
pytorch2.5.1	38	3
pytorch2.9	39	3
pytorch2.7	39	3
pytorch2.8	38	3

Additional details and impacted files

@@            Coverage Diff            @@
##           master   #21402     +/-   ##
=========================================
- Coverage      87%      79%     -8%     
=========================================
  Files         269      266      -3     
  Lines       23804    23772     -32     
=========================================
- Hits        20626    18730   -1896     
- Misses       3178     5042   +1864     

[justusschock]

Thanks @arrdel , could you update the changelogs? Other than that your PR seems fine to me :)

[arrdel]

Thanks @justusschock! I've updated the Fabric changelog as requested. The entry documents the fix for the DDP notebook CUDA fork check to allow passive initialization. The changelog is now complete and ready for review.