[BugFix] Evaluate MicroDuck head pose from the supplied state

[vmoens]

Summary

Fixes MicroDuck reward inspection for simulator snapshots that are not the backend current state.

Previously reward_features(state, action) derived the torso direction from state but the head gaze from the live backend site transform. A tilted snapshot could therefore receive the head pose of a different state.

Backend site rotation queries now optionally run forward kinematics for a supplied qpos without mutating live state. The normal reward hot path continues to use already-current derived data.

Tests

• Added a regression that evaluates a tilted snapshot while the live backend remains level and verifies both values independently.
• Native MuJoCo MicroDuck selection: 15 passed, 1 CUDA-only test skipped.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/4393

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit f80e340 with merge base 2d258fe ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.