When Preferences Diverge: Aligning Diffusion Models with Minority-Aware Adaptive DPO