Add nll_with_ignore and cross_entropy_with_ignore

[embedding-shapes]

Summary

Adds ignore_index support to cross-entropy and NLL loss functions via new *_with_ignore functions, enabling proper handling of padding tokens in sequence modeling.

Motivation

Sequence modeling (NLP, time series) requires excluding padding tokens from loss computation. Currently users must manually mask losses, which is error-prone and less efficient.

Changes

• Add loss::nll_with_ignore(inp, target, ignore_index) function
• Add loss::cross_entropy_with_ignore(inp, target, ignore_index) function
• Implement safe target sanitization to prevent OOB errors
• Add 4 tests including safety and PyTorch parity checks
• Fully backward compatible (new functions, existing ones unchanged)

PyTorch Compatibility

Matches PyTorch functionality with two documented divergences:

1. All-ignored case: Returns 0.0 (PyTorch returns nan) for better ergonomics
2. Type: Uses u32 (PyTorch uses i32) to match Candle's target dtype

Open Questions

• Is adding two new functions better for you rather than changing the signature of existing functions? I'm personally fine with both, opted for the approach that reduces API churn for now, but OK to change it if you want to.
• Are there too many comments? I opted for adding "too much" over "too little" right now, let me know if it's too verbose and you want to adjust. I'm still a beginner ML enthusiast, so the comment were more for helping me during the implementation, we might want to remove some of them if deemed "too basic" maybe.