feat: support KV cache CPU offloading in vLLM

[qinganrice]

Summary:
This PR adds support for using vLLM CPU offloading with KVCached. KV cache blocks are moved between GPU and CPU so that repeated prompts can be served from CPU cache instead of recomputing.

Changes:
KVConnectorMixinPatch

• Patches KVConnectorModelRunnerMixin.use_uniform_kv_cache() so it returns False when kvcached is enabled.
• Reason: With OffloadingConnector, vLLM’s prefer_cross_layer_blocks=True makes use_uniform_kv_cache() return True, so vLLM uses allocate_uniform_kv_caches() and a single cross-layer torch.zeros tensor. That skips kvcached’s _allocate_kv_cache_tensors path and the VMM elastic pool.
• By forcing use_uniform_kv_cache() to False, allocation goes through _allocate_kv_cache_tensors, so kvcached’s VMM-backed tensors are used and CPU offloading works.
• Applied only for vLLM >=0.12.0 via @version_range(VLLM_V12_RANGE).

Test:

• Runs successfully with vLLM 0.16.0.
• GPU→CPU offload, CPU→GPU fetch (cache hit) and CPU eviction (LRU & ARC) are tested.

[qinganrice]

For SGLang's support, KVCached requires --disable-radix-cache but HiCache (extended RadixCache for CPU offloading) requires radix cache to track GPU↔CPU load/store per prefix tree node and SGLang forbids using both flags together. Therefore, we cannot support SGLang CPU offloading so far.

[jiarong0907]

For SGLang's support, KVCached requires --disable-radix-cache but HiCache (extended RadixCache for CPU offloading) requires radix cache to track GPU↔CPU load/store per prefix tree node and SGLang forbids using both flags together. Therefore, we cannot support SGLang CPU offloading so far.

After support prefix cache, kvcached should no long require --disable-radix-cache.