Critique-Guided Distillation for Efficient and Robust Language Model Reasoning