fix(pd): support DeepSeek V4 grouped layerwise cache handoff

[lucifer1004]

Summary

• preserve DeepSeek V4 group-specific cache layouts and transfer manifests for P/D disaggregation
• publish and transfer typed cache segments as each semantic model layer becomes ready
• keep semantic layer topology independent from the number of physical cache slabs
• reject mixed prefill/decode batches at startup and at decode admission

This reworks and supersedes the closed P/D cache-handoff attempt in #660 against current main. The SM120 serving work remains independently scoped in #992.

Test plan

• pre-commit run --all-files
• P/D executor and manifest tests: 35 passed
• clean-main DeepSeek V4 configuration tests: 137 passed, 5 skipped, 2 unrelated SM120 tests deselected
• combined feat(deepseek-v4): support Flash serving on SM120 #992 + fix(pd): support DeepSeek V4 grouped layerwise cache handoff #997 focused suite: 179 passed, 1 skipped, 13 subtests passed
• InferLab P/D dry-run resolves enable_mixed_batch=false for both roles and omits the mixed-batch launch flag
• real four-GPU SM120 DeepSeek V4 Flash TP2/EP2 single-request P/D smoke with feat(deepseek-v4): support Flash serving on SM120 #992 applied: default layerwise interval transferred all 43 semantic layers on both TP ranks (86 completions), produced a coherent 16-token completion, and completed verified cleanup

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7d4ae46f89

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".