Text this: Greedy Prefetch for Reducing Off-Chip Memory Accesses in Convolutional Neural Network Inference