Surveying the Effects of Quality, Diversity, and Complexity in Synthetic Data From Large Language Models