Robustifying Learning-Augmented Caching Efficiently without Compromising 1-Consistency