Add model training, export, and read-only inference

[Kaushik-Kumar-CEG]

adds DeBERTa BIOES+CRF training, export, and read only inference.

the training command validates the generated dataset, selects the best checkpoint using validation F1, verifies the final model by reloading it offline, and writes a reproducibility manifest. Export supports CRF matrices and optional ONNX emissions. The inference API returns phrase text, offsets, model scores and truncation status without changing rules

Usage:

python -m pip install ".[training]"
train-required-phrase-model --data-dir dataset-output --output-dir model-output --model-revision COMMIT
export-required-phrase-model --model-dir model-output/final-model --output-dir model-export

Test it with:

./configure --dev
python -m pip install -e ".[training]"
venv/bin/pytest -q tests/test_training.py tests/test_model.py tests/test_export.py tests/test_inference.py
venv/bin/ruff check

The final model run reached 0.9008 validation F1 and 0.8766 test F1 with no invalid BIOES paths. This PR is stacked on PR #1