Change timing loss function to something depending on the evaluation time

[jacopok]

The current loss function in HyperparameterOptimization.objective doesn't make a lot of sense.

The validation error part is fine, but the other cost is currently in the form $CN + T$, where $C$ is some constant, $N$ is the number of training points used, while $T$ is the time required to train the network.

This should be replaced by some nonlinear function $f$ of the time $t$ required to evaluate the network. This could simply be evaluated by the number of connections....

[jacopok]

It would also be good to explore network complexity "sequentially" somehow, starting with smaller nets and going up as needed.