[jessiewei7]

Summary

• Add host-pool backup and restore for supported non-speculative
  DeepSeek-V4-Flash-0731 PD decode requests.
• Persist the SWA window used for backup and require matching FULL/SWA and
  DSV4 sidecar transfer lengths during restore.
• Abort only the affected request when backup or restore is unsupported.
• Allow cleanup after retraction before a new radix match, when no lock exists.

Unsupported configurations

Host-pool retraction is supported only for the non-speculative CUDA
DeepSeek-V4-Flash-0731 decode layout with a separate SWA pool and without
HiSparse.

DSPARK/MTP is not eligible for host-pool retraction. With auto selection it
uses the CPU-tensor fallback; if that cannot snapshot the request state, only
the affected request is aborted and the decode scheduler continues.

Validation

• 23 passed: focused DSV4 retraction and decode-queue tests.
• Regression coverage: cleanup before a new radix match; the parent code
  reproduces KeyError(None).
• 8×H20 1P1D PD run with forced retraction: four retractions completed with
  token IDs and text identical to the same-batch non-retracted control.
• Ruff, isort, Black, codespell, git diff --check, and repository pre-push
  checks passed.

Refs #33385

[hnyls2002]

Thanks @jessiewei7! The decode crash this targets is fixed on main by #43936: a retracted DeepSeek-V4 PD decode request is now sent back to prefill for recompute instead of raising NotImplementedError, and a request that cannot be recomputed (multimodal input, or the none backup backend) is aborted on its own while the scheduler keeps running. Keeping the KV on the host across retraction is still a valid optimization; that path is open in #37239. Closing as superseded; please reopen if I've missed a case.