[BUG FIX] Fix per-environment and PD control commands in the differentiable rigid-body simulation.

[duburcqa]

Description

Follow-up to #3087, fixing two ways differentiable-scene input commands were silently dropped, and closing the coverage gaps that hid them.

• The input tape keyed tracked setters by (method, dofs subset) only, so two same-step calls on distinct environment subsets evicted each other: the evicted environments' inputs received no gradient at all and were skipped by the backward replay. The key now includes the environment subset.
• The PD control setters (control_dofs_velocity, control_dofs_position, control_dofs_position_velocity) were neither tracked nor rejected in differentiable scenes, so per-step-varying targets were replayed as the last command for every step, corrupting the backward primals for all other tracked inputs. They are now tracked and replayed; a gradient-bearing PD target raises (use control_dofs_force for differentiable inputs).

Coverage: per-environment split inputs, per-step PD targets, multi-environment constraint tests (mixed per-environment limit activity), a tangential contact force component, and a nonzero cross-tree weld rotation bias. Both weld fixtures' relpose sign is also corrected (pose of body2 in body1's frame; the inverted value left them violated at rest - confirmed against MuJoCo).

Related Issue

Follow-up to #3087.

How Has This Been / Can This Be Tested?

• pytest tests/grad/ --backend cpu: 48 passed, 10 skipped on macOS M4 Max (fp64); 46 passed, 12 skipped on Apple Metal (fp32); 58/58 on linux CPU (fp64) and CUDA (fp32).
• The per-environment split arms and both PD-target arms were verified to FAIL on current main and pass with this PR.

Checklist:

• I read the CONTRIBUTING document.
• I followed the Submitting Code Changes section of CONTRIBUTING document.
• I tagged the title correctly (including BUG FIX/FEATURE/MISC/BREAKING)
• I updated the documentation accordingly or no change is needed.
• I tested my changes and added instructions on how to test it for reviewers.
• I have added tests to cover my changes.
• All new and existing tests passed.