Add more efficient validation function

[florian-huber]

Currently we use a generator for the training (which is good), but also for the validation step.
The later is not ideal as @niekdejonge pointed out.
We will only see one (or several if we want) fix batches, but that probably is not describing our complex data well enough.

It is probably better to add an efficient computation of an all-vs-all validation spectra comparison. This, together with a binned computation of mse/rmse.

[florian-huber]

Has been added with #179