RAD: Training an End-to-End Driving Policy via Large-Scale 3DGS-based Reinforcement Learning