[Feat] Integrate Mooncake into UCM By UcmMooncakeStoreV1

[NaganooMei]

Purpose

• Integrate UcmMooncakeStoreV1 into UCM/vLLM KV transfer flow, including lookup/load/dump/wait/check path support.
• Wire Mooncake v1 connector registration in store factory and pass runtime config from UCMConnector.
• Add Mooncake Prefix Cache user guide docs/source/user-guide/prefix-cache/mooncakestore.md.
• Update Prefix Cache docs index to include mooncakestore.

Modifications

• ucm/store/mooncakestore/mooncake_connector.py
  • Refactor and integrate Mooncake v1 backend logic.
  • Support async task execution and task lifecycle checks.
  • Add worker-side NPU buffer registration for RDMA transfer.
• ucm/integration/vllm/ucm_connector.py
  • Pass Mooncake-required runtime fields (device_id, tensor size metadata, optional register buffers, etc.).
• ucm/store/factory_v1.py
  • Register UcmMooncakeStoreV1 for connector factory creation.
• docs/source/user-guide/prefix-cache/mooncakestore.md
  • Add user guide for Mooncake Store v1 usage and launch instructions.
• docs/source/user-guide/prefix-cache/index.md
  • Add mooncakestore to toctree.

Test

Validated on single-machine Atlas A2 with whole-save whole-load enabled and UCM + Mooncake backend.

QwQ-32B (TP=2)

Sequence Length	Concurrency	Hit Rate	Baseline	Read	Improvement
8k	10	1.0	10819.43	1238.38	88.55%
8k	10	0.8	10819.43	3270.46	69.77%

DSV2-Lite (TP=2)

Sequence Length	Concurrency	Hit Rate	Baseline	Read	Improvement
8k	10	1.0	2911.66	859.18	70.49%
8k	10	0.8	2911.66	1356.10	53.43%