Underestimated Privacy Risks for Minority Populations in Large Language Model Unlearning