Near-Optimal Glimpse Sequences for Improved Hard Attention Neural Network Training