[codex] align latent factor training protocols

[stefan-jansen]

Summary

Aligns the library latent-factor training protocols with the Chapter 14 audit conclusions:

• adds CAE shuffled mini-batch training, hidden-layer BatchNorm, validation-loss best checkpointing, and per-member extraction
• makes SDF checkpoints phase-qualified, tracks validation-best states by phase, and preserves legacy integer checkpoint lookup where needed
• keeps IPCA KPS ThetaY normalization and final-gamma factor consistency cleanly documented without audit-era comments

Validation

• uv run ruff check src tests
• uv run ruff format --check src tests
• uv run ty check
• uv run pytest tests -q
• uv build

Book-side bridge validation was run against this local source via PYTHONPATH=/home/stefan/ml4t/libraries/ml4t-models/src because the current PyPI alpha predates these API changes.