On the Role of Structured Pruning for Neural Network Compression