Structured Labeling Enables Faster Vision-Language Models for End-to-End Autonomous Driving