feat: support shared-GPU simulation evaluation

[yufoo1]

Pin MuJoCo rendering independently of HTTP inference replicas, allowing GPU0 to render while already-running services on GPU0 and GPU1 both serve inference. Statically shard episodes across process-isolated workers and keep each worker's sessions on one replica.

Reuse the existing SimulationService lifecycle and original inference batcher. Add an evaluation CLI, paired placement configurations, result accounting, worker timeout and failure handling. Leave EmbodiInfer unchanged.

This integrates the deployment used by the earlier external-driver 1.762x warmed-loop result (1.485x including reset). This native runner has not been GPU-benchmarked.

Validation before removing the added test file from this commit: 1044 tests passed, 45 optional-dependency tests skipped; Ruff checks passed. Runtime code is unchanged by the documentation/test removal.

Summary

Type of change

• Bug fix
• New capability (robot, simulator, model, backend, or runtime feature)
• Documentation
• Refactor with no behaviour change
• Build, packaging, or CI

Verification

Paste the exact commands you ran and their result. State what you did not
verify (GPU, checkpoint, real hardware) rather than implying it works.

$ uv run pytest -q

Process boundaries

• client, deployment, application, devices, model_services
  remain the canonical domains, and services.* still imports.
• No inference-engine runtime code was imported into src/embodirun.
• Robot/simulator adapters and policy bindings stay separate.
• New robot- or model-specific dependencies are optional and declared in
  pyproject.toml.

Checklist

• I updated docs/en/support-matrix.md
  for any new combination, with versions, configuration, hardware,
  checkpoint, exact command, and observed result.
• I kept the English and Chinese READMEs in sync, if positioning, install
  steps, or the support matrix changed.
• No checkpoints, datasets, recordings, credentials, addresses, or personal
  data are included.
• I agree that this contribution is licensed under Apache-2.0, per
  CONTRIBUTING.md.

[hootandy321]

我看了下spec，并且在gpu服务器上跑了下应该没啥问题，让codex看了下代码这可能还有两个小点可以修改一下

GPU 小规模测试情况：

• 环境：Linux、NVIDIA L20，使用单卡。
• 模型与后端：π0.5 LIBERO 微调权重，固定版本 8e174154；原生 Pi05ServingAdapter 与 HTTP 服务，8 维状态输入、7 维动作输出，10 步去噪。
• Tokenizer：使用官方哈希一致的 SentencePiece 文件生成本地配置，242 项指令、状态边界及截断检查通过。
• 测试方式：对比 main 顺序执行、PR 单 worker 和双 worker。每组 2 个 episode、每个推进 3 步，共 18 次真实策略 HTTP 请求，使用受控采样随机性。
• 结果：三组短轨迹的状态、双相机图像哈希和实际动作逐字段一致，正常结束时 session 均成功释放。

[hootandy321]

1. 第 205–207 行：异常退出时的 session 清理
   超时或其他 worker 报错时，父进程会直接 terminate 子进程，可能跳过 finally 里的 DELETE。当前服务端没有 session TTL，遗留 session 会占住配额，累积后新请求就会返回 429，这个在 macOS 和 Linux 上都复现了。
   建议先让 worker 协作退出，再由父进程追踪并兜底清理 session。只补 finally 还不够，因为硬终止时它仍可能不执行。
2. 第 146 行：ID 加后缀后可能超长
   simulator_id 追加“-0”等 worker 后缀后，原本合法的 127／128 字符 ID 会超过接口的 128 字符上限，创建 session 返回 400。可以用 metadata 区分 worker，或者生成长度受限、不会冲突的派生 ID。
   麻烦给这两项补一下回归测试：超时、其他 worker 失败后的清理和配额恢复，以及 ID 长度边界。这两项修好、测试和 CI 通过后，就可以合并这版功能。