fix(sim/newton): an MJCF position servo keeps its damping and torque ceiling

[cagataycali]

Newton's MJCF importer reads a <position> actuator's kp into joint_target_ke and drops the rest of the servo: the dampratio MuJoCo compiles into a velocity gain (-actuator_biasprm[2]) and the forcerange that caps the torque. The joint arrives P-only with a 1e6 N m ceiling, so a held position command oscillates instead of settling. Both values are now read off the compiled model and written onto the builder before finalize.

send_action({"Rotation": 0.5}) on the shipped so100, held 3000 substeps

backend	kd	effort limit	final	max	overshoot
MuJoCo (reference)	compiled	3.5 N m	0.4997	0.4997	none
Newton, before	0.0	1e6 N m	0.2991	0.9559	91.2%, never settles
Newton, after	5.13	3.5 N m	0.4996	0.4996	none

Tests

tests/simulation/newton/test_mjcf_servo_gains.py (11 cases, MuJoCo-only - no Warp or GPU): the reader on dampratio / explicit kv / a plain motor, the writer landing on the DOF index rather than the joint ordinal behind a free base, and an unreadable model keeping the gains Newton did carry. tests/simulation/newton + docs scope 828 passed / 30 skipped; whole-tree graders 6183 passed / 77 skipped; ruff + mypy clean.

Serves 0.7 in #3818: a coverage matrix row is only worth publishing if the two backends mean the same thing by one command. +353 LOC, all of it the reader, its pins and the docs note.

[yinsong1986]

Summary

Newton's MJCF importer carries a <position> actuator's kp into joint_target_ke but drops the compiled velocity gain (-actuator_biasprm[2], from dampratio/kv) and the forcerange torque ceiling, leaving joints P-only with a 1e6 N m limit that oscillates instead of settling. This PR adds a small MuJoCo-backed reader (actuator_gains.mjcf_joint_servos) plus a writer (apply_joint_servos) that lands both values on the correct builder DOF indices before finalize, hooked into _rebuild for MJCF imports only. The hook runs under the _rebuild lock contract, only touches the two values Newton drops (kp is deliberately left to Newton's own importer), scopes writes to the just-imported robot via first_joint, and degrades gracefully (keep imported gains + log) when MuJoCo cannot compile the model — which is the pre-PR behaviour, so nothing regresses on that path.

What's good

• DOF-index (not joint-ordinal) writes, consistent with the existing joint_qd_start mapping convention in _rebuild, with a dedicated free-base regression test pinning the trap.
• The fallback path (unreadable MJCF keeps imported gains) is documented, narrow in effect, and regression-tested rather than asserted.
• Empirical cross-backend validation in the description (so100 step response, MuJoCo vs Newton before/after) rather than gains-in-isolation claims.
• Changelog fragment, docs note, ASCII-clean strings, no host paths in tests — AGENTS.md hygiene throughout.

Verification suggestions

• Spot-check the so100 step response on a Newton-capable box: send_action({"Rotation": 0.5}), hold ~3000 substeps, confirm settling near 0.4997 with no overshoot, matching the table in the description.