Fix TensorBoard logging compatibility with numpy >= 2.4.0

[paipeline]

What does this PR do?

Fixes #21503

This PR resolves TensorBoard logging issues with numpy >= 2.4.0 by properly handling numpy arrays and the new TypeError behavior when converting 0-dimensional arrays to scalars.

Changes Made

Bug Fix - TensorBoard Logger (Fabric)

• Enhanced log_metrics() method in src/lightning/fabric/loggers/tensorboard.py to properly handle numpy arrays
• Added fallback mechanism for numpy >= 2.4.0 TypeError when calling .item() on 0-dimensional arrays
• Supports both 0-dimensional and single-element 1D numpy arrays with appropriate scalar conversion
• Provides clear error messages for unsupported multi-dimensional arrays

Comprehensive Test Coverage

• New test file: tests/tests_fabric/loggers/test_tensorboard_numpy_240.py
• Tests cover: 0-dimensional arrays, 1D single-element arrays, multiple numpy dtypes, TypeError fallback simulation, invalid multi-dimensional arrays, mixed type logging, and backward compatibility

Backward Compatibility

• Zero breaking changes - all existing functionality preserved
• Torch tensors continue to work exactly as before
• Python scalars and dictionaries handled identically

Root Cause Analysis

Starting in numpy 2.4.0, converting 0-dimensional arrays to scalars raises a TypeError instead of silently succeeding. The TensorBoard logger's log_metrics() method only handled PyTorch tensors with .item() but didn't account for numpy arrays being passed in practice.

Solution Strategy

1. Array Detection: Check for objects with ndim and item attributes (numpy-like arrays)
2. Dimension Handling:
   • 0-dimensional arrays: Try .item(), fall back to float() conversion
   • 1-dimensional single-element: Try .item(), fall back to float(v.flat[0])
   • Multi-dimensional: Raise informative ValueError
3. Graceful Degradation: Handle TypeError from numpy >= 2.4.0 with appropriate fallbacks

Testing

The fix includes comprehensive tests that verify:

• 0-dimensional numpy arrays log correctly
• 1-dimensional single-element arrays work
• TypeError fallback for numpy >= 2.4.0 behavior
• Multi-dimensional arrays raise appropriate errors
• All existing tensor/scalar logging unchanged
• Mixed type logging in single call
• Multiple numpy dtypes supported

Impact

• Fixes: Critical logging failures for users with numpy >= 2.4.0
• Enables: Seamless numpy array logging in TensorBoard
• Maintains: Full backward compatibility
• Affects: Both lightning.fabric.loggers.TensorBoardLogger and lightning.pytorch.loggers.TensorBoardLogger (inheritance)

This ensures PyTorch Lightning users can safely upgrade to numpy >= 2.4.0 without breaking their TensorBoard logging workflows.

---

📚 Documentation preview 📚: https://pytorch-lightning--21544.org.readthedocs.build/en/21544/