[Refactor] Back OperationalSpaceController with Newton's operational-space controller

[hujc7]

Superseded by #7854, which consolidates the three supported controllers on the released Newton API. Ackermann remains a separate draft in #6654.

Summary

OperationalSpaceController now evaluates its control laws through Newton instead of an in-tree
Torch implementation: 2.6-2.8x faster per step (0.32 ms vs 0.88 ms at 4096 envs) with the
public configuration, command and output contracts unchanged.

Important

Blocked — do not merge or run CI yet. This depends on
newton-physics/newton#4105 (unmerged), plus
three further Newton changes described in §3 that are not yet proposed upstream. Against
develop's pinned newton[sim]==1.5.1, compute() raises
ImportError: cannot import name 'ControllerOperationalSpaceModelFree'. Opened to review the
Isaac Lab side of the change only.

Description

Fixes # (issue)

The task-space impedance law, the contact-wrench law and null-space control move to
newton.controllers.ControllerOperationalSpaceModelFree. Target resolution, the task frame and the
gain schedule stay in Isaac Lab, mirroring the JointImpedanceController port.

1. Frame handling

The task frame is handed to Newton as the operational frame, so gains, selection axes and targets
stay expressed in it instead of being rotated into the root frame first. That retires the
root-frame gain buffers, both selection matrices, the operational-space mass matrix and the cached
joint-space mass-matrix inverse, and lets the per-axis gains be stored as (num_envs, 6) vectors
rather than diagonal matrices. _motion_p_gains_task / _motion_d_gains_task change shape
accordingly; the one in-tree consumer (ExportPatcher) is updated.

2. Behavior

No intended behavior change. Verified two ways:

• Parity oracle — test_operational_space_newton_integration.py reimplements the previous Torch
  law and compares against it across 14 configurations (absolute/relative pose, open- and
  closed-loop wrench, all three decoupling modes, null-space centering, both variable impedance
  modes, with and without a task frame). 14/14 on cpu and on cuda:0.
• Existing sim suite — test_operational_space.py is 18/18, matching the pre-change baseline
  measured in the same container.

3. Required Newton changes

Prototyped locally and measured; not yet proposed upstream. Without them the sim suite is 12/18.

#	Change	Why
1	Adaptive Tikhonov ridge in _invert_spd_block_kernel	It floors the Cholesky pivot at _FLOAT32_EPS * max(|diag|, 1) then divides by that floor, so a near-singular J M⁻¹ Jᵀ inverts to order 1/eps. At cond = 2.4e8 the torque error is 349%, which failed 5 tests. Damping only on a second pass, and only when the first hits a degenerate pivot, keeps well-conditioned blocks bit-identical.
2	Gate the >=6 controlled-DOF guard on that ridge	Isaac Lab has always accepted inertial decoupling on under-actuated arms; Newton rejects it at construction. The ridge makes the rank-deficient case well-posed.
3	mask_motion_after_inertia flag	Newton applies Khatib's Ω to the commanded acceleration, ahead of Λ; this controller applies it to the resulting force, after Λ. The two differ only when a motion axis is de-selected and decoupling is on, but there the pre-Λ ordering misses the shipped contact-force thresholds (5.8 N and 7.4 N against a 5.0 N band).

4. Known contract narrowing

• compute() returns a view of Newton's output port rather than a freshly allocated tensor. Both
  in-tree callers copy or consume it immediately; external code holding the reference across steps
  would now see it mutate.
• Inputs must be float32; the previous Torch law accepted float64.
• The backend is built on the first compute() call, since num_DoF comes from the Jacobian, so
  configuration errors surface at the first step rather than at construction.

Type of change

• Breaking change (existing functionality will not work without user modification)

Release backport

• Backport this pull request to the active release branch after it merges into develop

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with ./isaaclab.sh --format
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package
• I have added my name to the CONTRIBUTORS.md or my name already exists there