[codex] Balance family training data

[henry3260]

Summary

• add a reusable deterministic helper that downsamples grouped fingerprints to the smallest non-empty family
• balance family data by default for train and build-templates
• expose --balance/--no-balance and --balance-seed CLI options
• report original and selected sample counts during training/template builds

Why

Uneven family counts can bias classifier training and family-template calibration. Using the same deterministic seed lets both workflows select the same balanced family-level dataset without deleting raw fingerprints.

Tests

• python -m pytest tests/training_data/test_training_data.py
• python -m pytest

Result: 43 passed, 3 existing datetime.utcnow() deprecation warnings.

Notes

• generated model/*.joblib artifacts are intentionally excluded
• model-version templates are intentionally unchanged because family-level balancing does not apply to that workflow