On Centralization and Unitization of Batch Normalization for Deep ReLU Neural Networks