Provable Filter Pruning for Efficient Neural Networks