Generating Training Data with Language Models: Towards Zero-Shot Language Understanding