fix(sim): make the camera-sensor tests pass reliably on CI (isaacsim / isaacgym / mujoco)

[juan-g-bonilla]

Problem

The camera-sensor test suites added in #170 failed across every backend lane. Root causes & fixes:

1. IsaacSim teardown deadlock (the 700s hang). On process exit, Kit's carbOnPluginShutdown deadlocks tearing down the omni.syntheticdata / OmniGraph render-product graph a TiledCamera creates (confirmed with a native py-spy stack). The subprocess never exits, so run_harness's timeout SIGKILLs it — turning a PASS (verdict already written to --result-file) into a spurious timeout. Fix: each camera *_assert.py os._exit(main()) after the sentinel write, bypassing the atexit teardown — the pattern behavior_assert / scene_spawn_assert already use (which is exactly why those isaacsim tests passed).

2. IsaacSim camera didn't follow a moving body / frame never refreshed (unmasked once the hang was fixed: follow, follow_multi_env, actor_mount, obs_pipeline). simulate_at_each_physics_step did sim.step(render=False) + a separate self.render(), but IsaacLab's self.render() only flushes fabric / drives the RTX render at render_mode >= PARTIAL_RENDERING; at NO_GUI_OR_RENDERING it is a no-op, so the TiledCamera returned its stale first frame. Fix: render inline via sim.step(render=render_now), which drives the low-level render regardless of render_mode (robust even if a caller reaches the sim before enable_cameras/headless are set at launch).

3. IsaacSim multi-env misalignment. IsaacSim ignores the passed env_origins and clones envs on its own env_spacing grid; all internal placement uses self.scene.env_origins, so sim.env_origins disagreed with where envs actually were (unlike mujoco/isaacgym, which honor the passed value) and the multi-env harness pinned robots away from their grid-placed panels. Fix: reconcile self.env_origins = self.scene.env_origins in create_envs so sim.env_origins means the same thing on every backend, and pin robots relative to it.

4. Un-actuated-robot drift flake (IsaacGym multi-env geometry/follow; a ~4px IsaacSim geometry offset). The spawned robot is un-actuated, so its root and joints drift during the settle and tilt the mounted camera. Fix: the camera harnesses now re-pin the root, hold the joints rigid at the spawn pose with zero velocity (set_dof_state_tensor_robots), read-back-verify the pose landed, and — since an IsaacSim pose write takes a few physics steps to reach the render — step a few frames before capture.

Also: a ruff PLC0415 per-file-ignore for tests/simulators/*_assert.py (these standalone harnesses defer torch/isaaclab imports until the SimApp is launched, inherent to IsaacSim).

Verification (run locally)

• isaacsim camera suite: 11/11 twice, plus follow at num_envs 1 and 4.
• isaacgym camera suite: 12/12, with follow_multi_env and geometry_multi_env each 6/6 across repeated runs (drift flake eliminated).