### What this PR does / why we need it? Mooncake Layerwise Connector supports hybrid attention manager with multiple kvcache groups. ### Does this PR introduce _any_ user-facing change? Yes. ### How was this patch tested? By CI. - vLLM version: v0.16.0 - vLLM main: https://github.com/vllm-project/vllm/commit/15d76f74e2fdb12a95ea00f0ca283acf6219a2b7 --------- Signed-off-by: nwpu-zxr <zhouxuerong2@huawei.com>