Clean up rigid robot backend code

[abcamiletto]

Summary

• require explicit backend namespaces in rigid core FK functions instead of optional inference paths
• reuse shared rigid hierarchy accumulation for BrainCo coupled-joint FK
• remove package-level to_qpos forwarding helpers that only called model.to_qpos

Claude review

Claude suggested several larger simplifications. I implemented the parts that keep the code minimal without a large redesign: removing optional xp fallbacks, unifying BrainCo joint_indices naming, and sharing the FK accumulation. I intentionally did not add a private to_qpos hook or collapse all backend/model wrapper layers in this cleanup PR.

Verification

• uv run ruff check src/body_models/common/rigid.py src/body_models/robots/g1 src/body_models/robots/brainco src/body_models/robots/smpl_humanoid src/body_models/skeletons/myofullbody tests/test_backend_consistency.py tests/test_smpl_humanoid.py
• uv run ruff format --check src/body_models/common/rigid.py src/body_models/robots/g1 src/body_models/robots/brainco src/body_models/robots/smpl_humanoid src/body_models/skeletons/myofullbody tests/test_backend_consistency.py tests/test_smpl_humanoid.py
• uv run --group dev --extra torch --extra jax ty check
• uv run --group dev --extra torch --extra jax -m pytest tests/test_backend_consistency.py tests/test_smpl_humanoid.py -q