feat: add number token loss implementation

[happybear-21]

feat: Add Number Token Loss (NTL) implementation

Issue: #38950

Add Number Token Loss (NTL) to improve language model performance on numerical tasks.
NTL addresses the fundamental limitation of cross-entropy loss on numerical tokens
by incorporating ordinal information into the training objective.

Key features:

• NTL-WAS: Uses Wasserstein-1 distance between numerical values
• NTL-MSE: Uses Mean Squared Error between numerical values
• Automatic numerical token detection (integers, floats, number words, ordinals)
• Efficient implementation with minimal computational overhead
• Caching of token-to-number mappings for performance
• Backward compatibility with fallback to standard cross-entropy

Implementation includes:

• Core NTL functions in src/transformers/loss/number_token_loss.py
• Integration with LOSS_MAPPING in loss_utils.py
• Comprehensive test suite in tests/test_number_token_loss.py
• Working example in examples/number_token_loss_example.py

The loss is designed to augment cross-entropy for tasks involving numerical
reasoning, mathematical operations, and any scenario where token ordering
matters for numerical values.

Resolves the issue where predicting "6" vs "9" for target "5" yields
the same cross-entropy loss, despite "6" being numerically closer.

[Rocketknight1]

Closing as we probably don't want this feature!