add merge batch normalization tools

[NHZlX]

When the training process is finished, we can merge the batch normalization with the convolution or fully connected layer. Merge batch normalization speeds up the forward process by around 30%.