feat(training): add standalone two-phase training

[heiwang1997]

Summary

• add a standalone two-phase training pipeline for calibrated NCore V4 data, including example configurations, supervision, losses, differentiable rendering, optimizer scheduling, and documentation
• add deterministic invalid-sample retry, distributed-safe experiment logging, and exact checkpoint/resume and preemption handling
• add regression coverage for normalization, seeded sampling, renderer tensor contracts and post-processing order, and masked RGB loss

Validation

• uv lock --check
• ruff check --no-cache .
• PYTHONDONTWRITEBYTECODE=1 .venv/bin/python -m pytest -p no:cacheprovider -q tests — 616 passed
• completed one local single-GPU context step and one rendering step

Scope

This change supports the dense front-camera two-phase path. Other model
families and variable-camera curricula remain unsupported. The GPU smoke test
covers one step only; it does not establish renderer-kernel bitwise
equivalence, distributed convergence, or final model quality.