Created by: Jie-Fang

1. fix bug of transforming tensor in BatchNormGradKernel::Compute, add x dims constraint. BatchNormKernel::Compute has added, but forgot to add in GradKernel.

2. fix bug in batchnorm unittest, change data_format to data_layout.

3. add more situations on amp unittest in order to cover batchnorm op test.