FLOPs as a Direct Optimization Objective for Learning Sparse Neural Networks