Improving Compositional Generalization with Self-Training for Data-to-Text Generation