Optimize Newton replicated camera startup

[kellyguo11]

Description

Optimize the full replicated-scene startup path used by the 4K Shadow Hand
camera task while preserving heterogeneous-environment support.

The change is intentionally kept in one draft PR so the cloning, camera,
Newton import, mesh/collision, articulation, sensor, and semantic-mapping
optimizations can be tested together.

What changed

• Construct Shadow Hand reorientation assets and cameras inside
  ReplicateSession, so backend replication sees the complete asset queue.
• Avoid automatic concrete asset clones for pure headless Newton + Newton Warp
  runs; retain concrete USD for PhysX, RTX, GUI/offscreen output, visualizers,
  and explicit USD clone consumers.
• Use compact internal USD references when concrete destinations are required.
• Resolve camera prototypes through the clone plan, cache shared camera
  intrinsics, and skip Newton-irrelevant physics validation in FrameView.
• Import clone-external Newton USD roots directly instead of traversing the
  replicated environment tree from the pseudo-root.
• Scope heightfield and collision-approximation scans to imported roots/shapes,
  and register only the selected solver's builder attributes.
• Cache repeated articulation target-mode resolution and reuse joint-wrench and
  site-frame view work.
• Resolve segmentation labels through the clone plan, reading semantic labels
  once per prototype while preserving per-environment instance identities and
  heterogeneous source selection.

Dependencies: none.

Performance

Measured against upstream/develop at e3c7dd87b6 on an RTX PRO 6000
Blackwell with warm asset and kernel caches.

Shadow Hand camera, 4096 environments

uv run --frozen isaaclab benchmark startup \
  --task Isaac-Reorient-Cube-Shadow-Camera-Benchmark-Direct \
  --num_envs 4096 \
  physics=newton_mjwarp renderer=newton_renderer \
  --benchmark_formatter summary

Metric	Before	After	Change
Scene creation	19.34 s	3.47 s	-15.87 s (-82.1%)
Builder finalization	2.52 s	2.46 s	-0.06 s (-2.4%)
Semantic mapping	5.14 s	0.62 s	-4.53 s (-88.0%)
Solver initialization	1.70 s	1.69 s	effectively unchanged
CUDA graph capture	1.51 s	1.52 s	effectively unchanged
Simulation start	14.82 s	8.46 s	-6.36 s (-42.9%)
Profiled environment creation	35.83 s	13.57 s	-22.26 s (-62.1%)
First step	0.22 s	0.23 s	effectively unchanged
Profiled launch through first step	38.52 s	16.15 s	-22.37 s (-58.1%)

A second post-fix run measured 15.91 s through the first step, showing the
expected small run-to-run variance. This does not yet meet the 10 s
end-to-end goal. The largest remaining fixed costs include builder finalization
(~2.46 s), solver initialization plus graph capture (~3.22 s), and the remaining
model/callback/reset work around them.

Kuka-Allegro lift camera, 4096 heterogeneous environments

uv run --frozen isaaclab benchmark startup \
  --task Isaac-Lift-KukaAllegro-Camera \
  --num_envs 4096 \
  physics=newton_mjwarp renderer=newton_renderer \
  presets=single_camera,rgb64 \
  --benchmark_formatter summary

Metric	Before	After	Change
Scene creation	13.13 s	12.42 s	-0.71 s (-5.4%)
Simulation start	8.73 s	7.35 s	-1.38 s (-15.9%)
Profiled environment creation	25.10 s	22.67 s	-2.43 s (-9.7%)
First step	0.53 s	0.52 s	-0.01 s (-1.9%)
Profiled launch through first step	29.39 s	26.93 s	-2.46 s (-8.4%)

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Performance improvement

Validation

• 33 focused tests passed, covering replication policy, compact USD references,
  articulation caching, scoped global/collision imports, clone-plan semantic
  mapping, and Kuka heterogeneous rendering.
• Verified the new regression tests fail against unmodified upstream/develop.
• Kuka heterogeneous Newton-renderer golden tests passed for both RGB and
  semantic segmentation.
• Isaac-Lift-KukaAllegro-Camera: RSL-RL, 16 environments, Newton MJWarp +
  Newton renderer, two learning iterations completed.
• Isaac-Reorient-Cube-Shadow-Camera-Direct: RSL-RL, 16 environments, Newton
  MJWarp + Newton renderer, two learning iterations completed.
• Both 4096-environment benchmarks above completed through the first step.
• uv run --frozen python tools/changelog/cli.py check develop: passed.
• uv run --frozen isaaclab -f: passed.
• Local Isaac Sim PhysX + RTX benchmarking is unavailable in this kitless
  worktree because the full Isaac Sim/Kit runtime is not installed. This remains
  an explicit CI/reviewer validation request.

Screenshots

Not applicable.

Checklist

• I have read and understood the contribution guidelines
• I have run the pre-commit checks
• Documentation changes are not required for this internal startup path
• 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 CONTRIBUTORS.md or my name already exists there