Less is More: A Stealthy and Efficient Adversarial Attack Method for DRL-based Autonomous Driving Policies