Learning Diverse Risk Preferences in Population-based Self-play