upgrade tensorflow code to NCHW format and fused bn

[suiyuan2009]

in my test NHWC format is 25% slow than NCHW on resnet101, and non fused batch normalization is very slow and uses much more memory than fused bn.