Zero-Shot Scene Understanding with Multimodal Large Language Models for Automated Vehicles