[vllm] Support speculative decoding in vllm rolling batch (#2413)

[vllm] Support speculative decoding in vllm rolling batch (#2413) #7328