Training-Free Restoration of Pruned Neural Networks