feat(model): add CRF layer for valid BIO sequence decoding

[hanneshapke]

Summary

• Add pytorch-crf dependency and CRF layer on top of PII token classifier
• CRF learns transition constraints between BIO labels during training
• Add Viterbi decode() method for optimal sequence prediction at inference
• Export CRF transition matrix as JSON alongside ONNX model for Go-side decoding
• Update trainer to use CRF negative log-likelihood loss for PII task

Motivation

Independent per-token classification allows illegal BIO sequences (e.g., I-EMAIL after B-PHONENUMBER). With 24 entity types and synthetic training data, these edge cases leak PII. The CRF enforces valid sequences globally and typically improves entity-level F1 by 1-3 points on BERT-based NER.

Test plan

• Train model and compare entity-level F1 against baseline without CRF
• Verify no invalid BIO sequences in predictions after Viterbi decoding
• Validate ONNX export still works (CRF is not exported; transitions are separate JSON)
• Benchmark inference latency overhead from Viterbi decoding

Closes #256