Towards Accurate UAV Image Perception: Guiding Vision-Language Models with Stronger Task Prompts