[P/D][Feature]显存调度优化，及时释放tensor

vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py

@@ -215,6 +215,9 @@ class P2pNcclConnector(KVConnectorBase_V1):
 
                 inject_kv_into_layer(kv_cache_layer, kv_cache,
                                      request.slot_mapping, request.request_id)
+                tensor = self.p2p_nccl_engine.recv_store.pop(request.request_id + "#" + layer_name, None)
+                if tensor is not None:
+                    del tensor
 
     def wait_for_layer_load(self, layer_name: str) -> None:
         """Blocking until the KV for a specific layer is loaded into vLLM's