Driving with LLMs: Fusing Object-Level Vector Modality for Explainable Autonomous Driving