What does this PR do?

Fixes a RuntimeError when using label_smoothing_factor > 0 with multi-label classification.

Previously, users would get a confusing error:
RuntimeError: gather(): Expected dtype int32/int64 for index

This PR adds a clear validation check in LabelSmoother that detects this incompatible combination and raises a helpful error message instead:

ValueError: Label smoothing is not compatible with multi-label classification.
Multi-label classification uses one-hot encoded labels (float tensors),
but label smoothing requires integer class indices.
Please set label_smoothing_factor=0.0 in your TrainingArguments.

Changes:

• Add validation in LabelSmoother.__call__() to detect multi-label classification
• Raise clear ValueError with actionable solution
• Add test case to prevent regression

Testing:

• ✅ All existing tests pass
• ✅ New test covers the fix
• ✅ Original reproduction case now shows clear error

Fixes #40258

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@zach-huggingface @SunMarc @qgallouedec - trainer related changes