[Warp] Read parity fixture poses through the warp accessor

[hujc7]

Addresses review comments on #6900.

• The pose-command parity fixture reached numpy via .torch.cpu().numpy(), routing warp buffers through a torch view and a host copy to get data warp hands back directly. Now uses .warp.numpy().
• UniformPoseCommand imports moved to the top of the file; they were local without a cycle to justify it.

ProxyArray.numpy() is not equivalent and was tried first: unknown attributes forward to the torch view, so it inherits torch's CPU-only restriction and raises on the cuda tensors these fixtures use.

Not included, raised on #6900 and better handled separately:

• Dropping the _wp suffix. term_dones_wp belongs to a family of six (time_outs_wp, dones_wp, terminated_wp, _truncated_wp, _terminated_wp, _scratch_term_mask_wp); renaming one leaves the package less consistent than it is now. A rename of the family, with deprecations, is its own change.
• wp.static for the resolved threshold flags, and 2D/tiled rewrites of the reward kernels. Both are optimizations that change how the kernels are defined and warrant their own benchmarking.

Release backport

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

Test plan:

• pytest source/isaaclab_experimental/test/ — 205 passed, 1 skipped
• uv run isaaclab -f clean

[isaaclab-review-bot]

Isaac Lab Review Bot

This test-only follow-up changes the pose-command parity fixture to materialize NumPy arrays directly from Warp buffers, moves command imports to module scope, clarifies the purpose of the checked-in Warp task-support set, and adds the required skip changelog fragment.

• Design and architecture: Using the Warp accessor avoids routing Warp-owned buffers through an intermediate Torch view while retaining the fixture’s NumPy reference calculations. The module-scope imports are consistent with ordinary dependency usage in this helper.
• API: No public symbols, configuration fields, defaults, exports, or documented runtime behavior change. The helper continues producing a UniformPoseCommand with the same command-buffer shape and device behavior for existing test consumers.
• Implementation: The four accessor substitutions preserve the array dimensions used by body indexing and the NumPy quaternion and position calculations; CUDA data is still materialized on the host as required by NumPy, but without the intermediate Torch conversion path. No production implementation path is modified.

No blocking issues. No inline issue met the actionable-evidence threshold; the assessment above records the review feedback.

Automated review; human maintainers own approval decisions.

[greptile-apps]

Greptile Summary

This PR updates pose-command parity fixtures to read CUDA-backed pose data directly through the Warp accessor and moves pose-command imports to module scope.

• Replaces Torch-view CPU conversions with direct wp.array.numpy() reads.
• Documents why the checked-in Warp-supported task set is retained as a regression pin.
• Adds an empty changelog skip marker for the review follow-up.

Confidence Score: 5/5

The PR appears safe to merge, with no actionable defects identified in the changed test infrastructure.

The direct Warp accessor returns the expected host-side pose shapes for the synchronized fixture buffers, and the import and documentation changes preserve existing test behavior.

Important Files Changed

Filename	Overview
source/isaaclab_experimental/test/envs/mdp/parity_helpers.py	Direct Warp-to-NumPy fixture reads preserve the expected pose shapes and synchronization behavior in the current callers; the import relocation revealed no concrete collection failure.
source/isaaclab_experimental/test/envs/test_frontend_cfg_conversion.py	Adds explanatory comments clarifying that the supported-task set guards against unnoticed Warp-support regressions.

_{Reviews (1): Last reviewed commit: "Say why the warp-supported set is pinned..." | Re-trigger Greptile}

[isaaclab-bot]

Backported to release/3.0.0 as 1344cfb.