Cooperative Meta-Learning with Gradient Augmentation