Learning a Universal Template for Few-shot Dataset Generalization