test: benchmark end-to-end SuperDex Go2 collector throughput

[TATP-233]

改动

新增短窗口 Go2 SuperDex collector benchmark，覆盖策略前向、动作处理、env.step、SuperDex physics、观测、reward、termination/reset 和 rollout bookkeeping。只测 collector 吞吐，不执行 PPO update 或完整训练。

复现

PYTHONPATH=/path/to/UniLab/src:/path/to/unisim/src:/path/to/unilab_rl/src:/path/to/project_superdex/.venv/lib/python3.12/site-packages \
SUPERDEX_ASSETS_PATH=/path/to/project_superdex/assets \
uv run --no-sync --python /path/to/project_superdex/.venv/bin/python --with superdex-physics \
  scripts/benchmark/superdex_go2_collector.py --mode batched \
  --num-envs 128 --workers 16 --warmup 2 --steps 10 --repeats 2

本机结果

AMD Ryzen 9 9950X3D2，16 physical / 32 logical CPUs，Linux x86_64，single NUMA；2 warmup + 10 measured，2 repeats：

• unbatched：[21484.85599714566, 20846.778222819292]，median 21165.817109982476 collector-env-step/s
• batched：[21553.14268869044, 22604.2549436869]，median 22078.69881618867 collector-env-step/s
• 增益：+4.32%
• mean reward/step：0.01453 -> 0.01429，短窗口仅作完整性检查，不代表训练质量。

[TATP-233]

Superseded by consolidated release PR #1545, which includes the collector benchmark.