[HiCache] Make LayerSplit MLA cache backup CP-aware

[dblate]

Motivation

Make GLM-5.2 LayerSplit context-parallel cache backup compatible with page-first HiCache storage and Mooncake L3. CP ranks own distinct token shards, so each rank must publish its cache under rank-scoped keys.

Depends on #37231 for the per-layer LF-to-PF transfer operator.

Modifications

• Keep MLA storage backup enabled on every CP rank while preserving the single-writer behavior without CP.
• Add CP rank/size suffixes to Mooncake object keys and group IDs.
• Use the per-layer LF-to-PF fallback for layer-sharded MLA and DSA page-first write-back.
• Add CPU unit coverage for backup selection, Mooncake key isolation, and fallback dispatch.

Accuracy Tests

• H20, CUDA 12.9: LayerSplit MLA page-first fallback -> 1 passed.
• H20, CUDA 12.9: DSA indexer page-first fallback -> 1 passed.
• H20 development environment: CP backup, Mooncake key isolation, and HiCache dispatch -> 24 passed, 1 skipped.
• git diff --check and Python AST parsing passed for changed files.

Speed Tests and Profiling

No end-to-end serving benchmark was run for this runtime-only change. The underlying LF-to-PF kernel microbenchmark is reported in #37231.

Checklist

• Format code with pre-commit.
• Add unit tests according to the contribution guide.
• Update documentation according to the contribution guide.
• Provide targeted NVIDIA GPU validation.
• Run GLM-5.2 CP + HiCache + Mooncake end-to-end validation.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33382072906
Latest PR Test (Extra): ❌ Run #33382072655
Latest PR Test (AMD ROCm 7.2): ❌ Run #33382072894

[dblate]

/tag-and-rerun-ci

[dblate]

补充验证结果：在 H20（CUDA 12.9）开发空间运行 LayerSplit MLA page-first fallback，结果为 1 passed；DSA indexer fallback，结果为 1 passed。CP backup、Mooncake key isolation 和 HiCache dispatch 定向单测合计 24 passed, 1 skipped。

当前 CI 仍未开始真实测试，原因是缺少上游要求的 run-ci 标签；请维护者按贡献指南执行 /tag-and-rerun-ci。