[Core] Use zeromq to put request output tokens back to the api server

[s5u13b]

Previously, we use ray.queue as the request output queue to store the streaming request output tokens returned by llm engine. However, we found that the cost of ray rpc will become very high when the frequency of remote call is high.

Because the request output queue of a api server could be written by multiple llm engines due to dynamic serving of Llumnix, it means that the frequency of streaming output rpc can be very high. So we decide to use zeromq as the default rpc way of putting streaming request output tokens back to the request output queue.

Besides, rpc(tcp://...) has no more obvious performance interference with step than ipc(ipc://...).