RAC3: Retrieval-Augmented Corner Case Comprehension for Autonomous Driving with Vision-Language Models