[Model] Qwen-2-VL Support

[nihalgeorge01]

This PR adds support for the Qwen-2-VL (vision-language) model

[buqimaolvshangxue]

qwen2_vl can work by this commit ？ @nihalgeorge01 ， i also has the needs to support qwen2_vl

[nihalgeorge01]

Not yet, we are fixing some bugs in the code locally. Working on pushing this out soon

[buqimaolvshangxue]

Thank you very much for your work! When I was thinking about this problem, I found that when processing the llava model in mlc, the text embedding and the image embedding are directly spliced ​​together to get the final embedding. But it seems that in the approach of qwen2_vl in vllm, the image embedding replaces certain specific positions in the expanded embedding. I wonder if the direct splicing method of llava is feasible? But if the splicing embedding method is not adopted, it seems that the public interface function needs to be modified. @nihalgeorge01