Switch BatchNorm implementation to rely on the official tf.layers.batch_normalization

[tvercaut]

Batch Normalisation is officially supported in the main TensorFlow repo: https://www.tensorflow.org/api_docs/python/tf/layers/batch_normalization

It would be beneficial to rely on this rather than our own implementation: https://github.com/NifTK/NiftyNet/blob/dev/niftynet/layer/bn.py

Note that other users have done that already, see e.g.: https://github.com/gift-surg/fetal_brain_seg/blob/master/NiftyNet/niftynet/layer/bn.py vs. https://github.com/gift-surg/fetal_brain_seg/blob/master/NiftyNet/niftynet/layer/bn_backup.py

[tvercaut]

This relates also to #280