refactor(srt): bind script-installed component versions from the master config

[adibarra]

Summary

Component versions that repo scripts install now come from the master config, not hardcoded pins.

• Scripts: these read their version from env and check it with check_env_vars, with no defaults:
  • vllm-router.sh, renamed from vllm-router-0.1.14.sh; reads ROUTER_VERSION.
  • vllm-mooncake.sh, renamed from vllm-mooncake-0.3.11.sh and now also called by kimik3-b300-mooncake.sh and the mooncake-master preambles; reads KV_OFFLOAD_BACKEND_VERSION.
  • lmcache-mp-rocm.sh; reads KV_OFFLOAD_BACKEND_VERSION.
  • glm5.3-tilert-rocm.sh; reads ROUTER_VERSION, replacing the recipe's TILERT_VERSION.
  • All use set -eo pipefail.
• Binder: writes the master router / kv-offload-backend version into each place the install runs: top-level environment for setup scripts, and services[].env for preambles that run an installer.
  • A missing master version, a mismatched component, or a fragment that sets these keys fails before submission.
  • The MI355X frontend's SETUP_PIP_PACKAGES vllm-router== pin is left in place and checked against the master version.
• Fingerprint: includes the bound versions, which come from the row, so a version bump changes the fingerprint and old results can't be reused.
• setup_script check: a setup_script that resolves to no file in configs/ (ours or the submodule's) now fails at launch and in infx generate. Upstream srtctl only warns.
• Docs: CONFIGS.md and RECIPES.md updated.

Verification

• Harness: every srt row × reachable cluster. The only differences are the intended ones: renamed setup_script in 74 recipes, ROUTER_VERSION added in 70, KV_OFFLOAD_BACKEND_VERSION added in 42, mooncake preamble changed in 39, and GLM-5.3 TILERT_VERSION replaced by ROUTER_VERSION.
• Installed versions: all 290 install sites, rendered with a stub python3, run identical pip commands base vs head except:
  • the H100 mooncake-master install now passes --quiet;
  • mooncake-master preambles now also run the MooncakeDistributedStore import check that worker installs already run.
• Tests: version bound into setup and service env; missing master version fails; each installer fails without its env var and installs exactly the bound version.
• Suite: the full CI pytest command passes 2319 tests, and ruff is clean.

No hardware smoke yet. One MI300X LMCache point and one Kimi-K3 B300 point would cover the script and service paths.

[github-actions]

Thanks for the contribution!

• Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
• PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
• After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.

中文

感谢你的贡献！

• **审阅：**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER，请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，再通过 Slack 联系核心维护者。必须严格遵循模板，并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have，才能触发签核验证。
• **PR 验证：**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目，并且恰好带有一个主标签时，才会运行扫描：full-sweep-fail-fast（强烈推荐；canary 加逐矩阵 fail-fast）、full-sweep-enabled（有 canary；矩阵任务在失败后继续运行）或 non-canary-full-sweep-enabled（无 canary，也无 fail-fast）。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
• **合并后：**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复；参见重新运行失败任务的说明。