Compressing Deep Reinforcement Learning Networks with a Dynamic Structured Pruning Method for Autonomous Driving