t2m-v2 Slice B: canonical normalization + template library v5

[fernandotonon]

Parent epic: #837, depends on Slice A. Rig→canonical extraction for scraped FBX/glTF clips (reverse of applyMotionClip; the BVH path in build-motion-library.py is the reference), active-window selection, tag→keyword vocabulary, library build → qtmesh-motion-library-v5 with hundreds of actions / several takes each. Ships as a template-library upgrade (default path) before any model training.

🤖 Generated with Claude Code

[fernandotonon]

Shipped in PR #843: canonical clip extraction (--dump-canonical) with per-clip restWorld/restDir measured at the clip's calmest animated frame, the bind-referenced direction-matching retarget (template + model paths), the v5 builder propagating the reference triple, and the eval harness. Library v5 (114 clips / 23 actions, all bind-referenced) is live on the HF models repo under motion/. Follow-ups: #854 (arm-space post-processing) and retraining the t2m model on the v5 data.