VideoRepair: Improving Text-to-Video Generation via Misalignment Evaluation and Localized Refinement