Negating Negatives: Alignment without Human Positive Samples via Distributional Dispreference Optimization