Optimal Quantization for Batch Normalization in Neural Network Deployments and Beyond