Reducing Non-IID Effects in Federated Autonomous Driving with Contrastive Divergence Loss