t2m v8.0: template-taught model with the leg-chain fix (#837)

Trained on the curated template clips (--library-repeat 40) rather than the
CMU corpus, which never carried the correct facing convention. Adds an
anatomical ankle gate and a leg-chain loss that keeps a walk's bend in the
KNEE rather than the ankle.

walk knee defect 0.359 -> 0.0003 (real Mixamo walk 0.014)
walk refDist 0.287 -> 0.243 (real-vs-real floor 0.437)
run refDist 0.472 -> 0.324

Vocab 34 -> 30: march/throw/hey/confession removed (broken source clips,
ankle folded ~100+ deg); those prompts fall back to the curated templates.
Previous model preserved as motion/t2m-v61.onnx.

motion/t2m.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:126ae26f2fdef4758a1ebe8a611b2ee83faee1f5366736de3c3ebba31583a365
-size 96564097
+oid sha256:eb243c917ecd775e9be28d4aa58ed32c6c17756ec6d33de1d894e9d26db00a4c
+size 96577959