Add local AI chart engine

[Chloiris]

What

• add the local AI chart engine: SM parsing, dataset building, optional Transformer training/inference, deterministic generation, footwork optimization, validation, and SM export
• add the reference-library and project chart workspaces with synchronized five-lane playback, selectable difficulty, 4x default visual speed, regeneration refresh, and validation feedback
• harden publication safety by excluding local corpora, audio, models, databases, exports, and private acceptance tests; replace private-derived fixtures with synthetic data; return relative reference paths from the API

Why

BeatForge had acoustic analysis and export workflows but no end-to-end local five-lane chart generation and preview flow.

Impact

The new workflow runs locally and requires the developer to provide a licensed reference corpus. No corpus media, user project data, model weights, or generated charts are included in this PR.

Root cause

Chart generation, playability validation, reference playback, and local-model integration were previously missing from the application boundary.

Checks

• .venv/bin/python scripts/beatforge.py test — 200 backend passed, 1 skipped; 70 frontend passed
• .venv/bin/python scripts/beatforge.py lint
• .venv/bin/python scripts/beatforge.py build
• staged secret/path scan and ignore-rule verification