Diverse, not Short: A Length-Controlled Data Selection Strategy for Improving Response Diversity of Language Models