Remove bias in augmented loss + fix perplexity

[drauh]

• removing bias in augmented loss calculation (needed to add a Bias layer for the weight tying model)
• now the last layer of the augmented model is a softmax. We use the log to recover pre-activation
• fix categorical crossentropy calculation for perplexity and augmented loss, it was the opposite ordering
• fixing running average perplexity calculation by using a hacky callback which exponentiated a running average of the cross-entropy. This replicate this implementation here (however because of the convexity this will only make its value bigger)
• removing explicit tensorflow dependency

Everything is running well but I couldn't check the performance for more than 2 epochs since I don't have a GPU so that training is very slow...