Learning Sparsity and Quantization Jointly and Automatically for Neural Network Compression via Constrained Optimization