Add Pages with Annotator Disagreement to Training Data

[carschno]

The current model is trained using non-empty pages on which both annotators agreed.
Include pages with annotator disagreement into the training data (with weighted and/or average score) could improve the results.

• Adapt training data selection method
• Retrain model
• Evaluate new model, and compare to current results