Maintains moving averages of all trainable variables

[wyli]

For many models trained with stochastic gradient descent, it is often beneficial to maintain a moving average of each model variable as the model evolves over training iterations. At inference time, the moving average version of the model is used instead of the model precisely at the last training iteration. This often gives better inference performance and should be implemented in the application driver (This will likely break the current variable restoring functions).

--originally posted at https://cmiclab.cs.ucl.ac.uk/CMIC/NiftyNet/issues/141