Mixed-Precision Neural Network Quantization via Learned Layer-wise Importance