Lost in Pruning: The Effects of Pruning Neural Networks beyond Test Accuracy