Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond 08-25