[vllm] Support speculative decoding in vllm rolling batch

[vllm] Support speculative decoding in vllm rolling batch #7326