Add GPU Mooncake layerwise connector

[Pz1116]

Summary

• add a GPU-safe MooncakeLayerwiseConnector without torch_npu or torch.npu dependencies
• support layerwise KV push with PP-aware side-channel ports and one prefill rank pushing token/head slices to multiple decode ranks for Prefill TP < Decode TP
• register the connector as MooncakeLayerwiseConnector in the v1 KV connector factory

Validation

• python -m py_compile vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_layerwise_connector.py vllm/distributed/kv_transfer/kv_connector/factory.py
• env XDG_CACHE_HOME=/tmp PRE_COMMIT_HOME=/tmp/pre-commit pre-commit run --files vllm/distributed/kv_transfer/kv_connector/v1/mooncake/mooncake_layerwise_connector.py vllm/distributed/kv_transfer/kv_connector/factory.py
• git diff --check