VLM-RL: A Unified Vision Language Models and Reinforcement Learning Framework for Safe Autonomous Driving