Rethinking Driving World Model as Synthetic Data Generator for Perception Tasks