Large Language Models are Few-Shot Training Example Generators: A Case Study in Fallacy Recognition