Enhancing Vision-Language Models for Autonomous Driving through Task-Specific Prompting and Spatial Reasoning