changes to allow training with validataion data

[dribnet]

This is a work in progress to allow training with a validation set. As written, the validation data set is required, but I would like to make it optional with a command line switch. I'm posting this as a pull request to solicit feedback before continuing.

Using a local set of MNIST image files, I can launch this via:

fauxtograph train --gpu --shape 28 28 --color_channels 1 \
  --epoch 100 --batch 100 --kl_ratio 0.1 --save_freq 10 \
  /data/mnist/png/mnist_train \
  /data/mnist/png/mnist_test \
  ./models/mnist_split1

The output to the console is:

epoch: 1
train_total_loss=0.0686056688428, train_rec_loss = 0.0683446452022, train_kl_loss = 0.00260960473679                                                                   
validation_total_loss=0.0668101087213, validation_rec_loss = 0.0655997917056, validation_kl_loss = 0.0121031077579                                                     
epoch: 2
train_total_loss=0.0662013068795, train_rec_loss = 0.0648920089006, train_kl_loss = 0.0130931474268                                                                    
validation_total_loss=0.0661751106381, validation_rec_loss = 0.0648957043886, validation_kl_loss = 0.0127941668034                                                     
epoch: 3
train_total_loss=0.0652597472072, train_rec_loss = 0.0632267221808, train_kl_loss = 0.0203300975263                                                                    
validation_total_loss=0.0652015134692, validation_rec_loss = 0.0633098185062, validation_kl_loss = 0.0189169626683                                                     
...

This can be cleaned up further, but wanted to first start a discussion about whether this type of functionality would be a worthwhile addition to the codebase.