Extend AIBrix kv cache to omni scenario

### 🚀 Feature Description and Motivation

https://docs.vllm.ai/projects/vllm-omni/ has its own connector base

```

OmniConnectorBase (abstract base)
├── SharedMemoryConnector — single-node, uses POSIX shared memory
├── MooncakeConnector — multi-node, uses Mooncake distributed store
└── YuanrongConnector — multi-node, uses Datasystem KV client
```

Data Flow (KV Cache Transfer)
```
The full flow is managed by OmniKVTransferManager (kv_transfer_manager.py):

Stage 0 (Thinker/LLM) Stage 1 (Talker/Diffusion)
───────────────────── ──────────────────────────
1. Prefill completes
2. _extract_kv_cache()
GPU blocks → CPU tensors
3. connector.put(kv_data)
→ serialize → mooncake.put()
4. connector.get(kv_data)
→ mooncake.get() → deserialize
5. apply_kv_cache_to_request()
CPU tensors → GPU

```

The AIBrix connector only needs to handle storing and retrieving byte blobs by key

### Use Case

Enable the kv cache transfer for omni distributed pipeline

### Proposed Solution

_No response_