Improving CLIP Robustness with Knowledge Distillation and Self-Training