Add per-head event weights for multi-head classification

[AvencastF]

Summary

Adds first-class independent event-weight support for multi-head classification.

• Accepts head -> Tensor[N] weight mappings for training, validation, and evaluation.
• Preserves shared single-tensor weights and the existing single-head API.
• Applies each head’s weights to its unreduced event losses, normalized by active weight sum.
• Excludes ignore_index (-100) rows from losses and metrics per head.
• Rejects sampler="weighted" with per-head mappings because no sampling-combination policy is defined.
• Adds validation, batching, evaluation/export support, CPU-focused tests, and documentation.

Why

Previously multi-head labels supported mappings, but weight mappings failed when the dataset attempted to convert the dictionary into a tensor.

Validation

• Static compilation and whitespace checks passed.
• Runtime tests require a Python environment with PyTorch; the available environment did not include it.

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.