Tweaks to EdgeClassifier to make Examples run properly

There were a couple of small issues that remained from the previous MR, which should be resolved here, namely:

• BCE is much harder to train than BCE_with_logits (therefore have switched back to having models produce raw outputs, then applying sigmoid only for evaluation)
• batch.scores is not set during inference, leading to an error in evaluation
• dataset_class is not in the example configs, leading to an error of hparams['dataset_class']. Fixed with hparams.get...