[codex] Fix SOMA prepared skinning weights

[abcamiletto]

Summary

• Fix SOMA prepare_skinning() so renderer payload weights match skinning_transforms.
• Add a named SOMA internal skinning-weight helper used by forward skinning and prepared payloads.
• Cover the public-vs-internal rig distinction with a regression test across NumPy, Torch, and JAX.

Root Cause

New SOMA assets expose a 77-joint public rig, while skinning uses a 121-joint internal rig. The inherited prepare_skinning() packed public model.skin_weights with internal skinning_transforms, producing incompatible renderer inputs.

Validation

• uv run pytest: 110 passed, 2 skipped
• uvx ruff check .
• uvx ruff format --check .
• uv run --group dev --extra torch --extra jax ty check
• uv build
• Manual SOMA checks against both local asset roots:
• /Users/abcamiletto/body-models-assets/models_hub/soma
• /Users/abcamiletto/body-models-assets/legacy_hub/soma