fix score if batch_iterator_test doesn't yield same order / size as y

[cancan101]

Closes #269

[BenjaminBossan]

I believe we should avoid having methods that return different things (single item vs tuple) depending on how you call them.

An alternative could be that we implement a method that transforms X and y, e.g.:

Xt, yt = net.batch_iterator_test.transform(X, y)
score = accuracy_score(y_true, yt)

This would also be useful in case someone needs the transformed X.

[cancan101]

This is something that should happen batch wise. What you have will pull all of the "transformed" Xs into memory.

[cancan101]

This is not just about transforming the y's (which is important), but also about the batch nature of access to the underlying data.