Add DeBERTa training and ONNX export pipeline for required phrases extraction

[Kaushik-Kumar-CEG]

part of #5137. Adds the training and export pipeline for the required phrase
NER model, building on the dataset from #5105

the model tags which words in a rule are required phrases, so we can predict
{{ }} markers for license expressions that have no coverage yet

this PR:

• train_model.py - loads the BIOES JSONL, aligns word labels to subwords,
  runs training with per epoch eval and entity F1. Small CLI (epochs, lrs,
  --limit, --resume)
• phrase_model.py - DeBERTa-v3-large with a word level CRF head and layer
  wise lr decay
• export_onnx.py - exports the emissions to ONNX and the CRF transitions
  beside it, with a numpy viterbi decoder for CPU inference
• test_train_model.py - unit tests for the alignment, metrics and viterbi
• etc/requirements-ml.txt - training only deps, separate from the runtime

update: first full training run is done

Test results on unseen licenses:

• F1: 0.822
• precision: 0.811
• recall: 0.833

This is a first run and already improves on my earlier BIO prototype (F1 ~0.79)
I expect later iterations (checkpoint averaging, more tuning, maybe adding negative examples) to do better

model: https://huggingface.co/Kaushik-Kumar-CEG/scancode-required-phrases-deberta-bioes-crf

disclosure : used Claude to help review and clean up bugs in script