Mind the Generative Details: Direct Localized Detail Preference Optimization for Video Diffusion Models