1. support channel last in BatchNorm*d (#27875)
2. fix a bug in batch_norm_op cuda kernel by extracting ResizeToChannelFist(Last), TransToChannelFirst(Last) to operators/layer_utils.h