Echoes Over Time: Unlocking Length Generalization in Video-to-Audio Generation Models