Optimized Support Sets for Few-Shot Learning