Saccader: Improving Accuracy of Hard Attention Models for Vision