Fine-Grained Evaluation of Large Vision-Language Models in Autonomous Driving