Distilling Out-of-Distribution Robustness from Vision-Language Foundation Models