build: pin UniSim 1.5.1 and compare MuJoCo collector throughput

[TATP-233]

Pin UniLab's normal and ROCm dependency profiles to released unisim-core==1.5.1, including the SuperDex extra requirement. Both locks continue to use PyPI. Align the inherited scene model-file default and update existing regression fixtures for the current backend implementation and native camera contract; no collector hot-loop change is introduced.

The requested 1.4.0 versus 1.5.1 MuJoCo collector comparison shows lower throughput in 1.5.1. On the same UniLab 84a00d04 code, Go2 flat task, 8 physical CPU cores, 30 warmup + 300 measured vector steps, and three fresh-process repetitions per version:

Environments	1.4.0 median env transitions/s	1.5.1 median env transitions/s	Change
1024	138,780	88,167	-36.47%
4096	209,773	120,476	-42.57%

The largest increase is in state update (N4096: 2.482 → 15.994 ms/vector step), while physics is 16.344 → 16.811 ms. All measured reset counts are zero. New final-body-state freshness work is a source-supported candidate explanation; this experiment did not disable that correctness fix or establish a per-function causal ablation. No non-MuJoCo throughput benchmark was run. G1 was excluded because unmodified 1.4.0 fails during mesh path resolution before timing.

scripts/benchmark/reports/unisim_1_4_0_vs_1_5_1_collector.md contains the configuration, ranges, paired changes, interpretation limits and reproduction commands. Its JSON and compressed raw artifact preserve exact package versions, source/asset hashes, commands and all 12 runs. The existing active-window benchmark includes environment, terminal handling/replay and bookkeeping; it excludes policy inference, learner waiting, process startup and cross-process IPC. The earlier aborted 1.5.0 comparison is excluded.

Validation on a9b8244e:

• uv sync --locked --extra mujoco --extra motrix: installed PyPI 1.5.1. The Motrix extra is only for the repository's existing validation requirements, not an additional throughput benchmark.
• make test-all: 1468 passed, 26 skipped, 588 slow deselected; 70% coverage; Ruff/mypy/Pyright passed (one optional Drake warning); benchmark entrypoint smoke 34/34 module and 35/35 script.
• uv lock --check: passed for standard and isolated ROCm metadata/lock pairs. No ROCm hardware benchmark claimed.
• git diff --check: passed.
• All six current-head CI jobs passed on a9b8244e842ccdee63a7053ccaf8fc46429288af, and Sphinx build passed. PR deployment is intentionally skipped.

Base: main. Related dependency/consumer tracking: #1599; package boundary: ADR-0007. This PR updates the dependency independently of the pending M2 consumer PR #1600. It does not merge the M2 consumer implementation or change training/collector lifecycle, tasks, rewards or algorithms.