feat:Optimize qwen2-vl to reduce cudaMemcpyAsync

[cynthieye]

qwen2-vl logic optimization: During each forward propagation, the xformer branch of Qwen2VisionTransformer will execute multiple tensor tolist methods (flash attn branch will execute multiple tensor items) to force the GPU tensor to be copied to the CPU, triggering CUDAMemcpyAsync to increase time consumption. Since the input and output are the same multiple times, it will be executed once, and the remaining will reuse the first result

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing/overview.html