Entropic Distribution Matching in Supervised Fine-tuning of LLMs: Less Overfitting and Better Diversity