feat(model): remove coreference resolution task

[hanneshapke]

Summary

• Remove coreference resolution task from the entire training pipeline
• Rename MultiTaskPIIDetectionModel → PIIDetectionModel, MultiTaskTrainer → PIIModelTrainer
• Remove MultiTaskLoss, coref classifier, coref loss weights, coref metrics
• Simplify preprocessing to skip coreference tokenization
• Remove create_coreference_sample from tokenization.py
• Update ONNX export to only output pii_logits
• Update eval, comparison, and pipeline scripts for single-task model

Motivation

The coref task competes for model capacity with noisy synthetic supervision and untuned loss weights. Removing it dedicates 100% of encoder gradients to PII detection — the primary and most critical task. This also simplifies the codebase significantly (-693 lines).

Test plan

• Train model and verify PII F1 is equal or better than multi-task baseline
• Verify ONNX export produces correct single-output model
• Run eval_model.py and compare_models.py to ensure they work with new model
• Verify Go proxy inference returns empty coref dict for backward compatibility

Closes #259