feat(turbomind): integrate LMCache for KV cache lookup, store, and retrieve

[irexyc]

Summary

• Add optional LMCache multiprocess-server integration through --lmcache-addr and TurbomindEngineConfig.lmcache_addr.
• Implement asynchronous LOOKUP, chunk-based STORE, and RETRIEVE using ZeroMQ, MessagePack, and CUDA IPC.
• Support attention KV caches and GDN checkpoints through unified cache layout planning and chunk-aligned checkpoint capture.
• Integrate retrieval with native prefix caching and scheduler allocation, preserving local cache contents and coordinating completion across TP ranks.
• Use RAII for transfer pins and lookup-lock ownership, with cancellation, failure fallback, and shutdown draining.
• Reuse the existing TP communicator and keep cache coordination encapsulated in the LMCache component.
• Exclude request modes requiring full prompt computation and incompatible dynamic-RoPE cache reuse.

[lvhan028]

model_path is passed to lmcache as model_name. Is there a better way?