Greedy-layer pruning: Speeding up transformer models for natural language processing