exclude spans that are out of valid window

[ArneBinder]

When doing inference with a TransformerTokenClassificationTaskModule that was trained with windowing and an overlap, it will also produce spans from the output of the model that is just the additional context, i.e. the part that overlaps. However, this is not recommended since the model was not trained to produce correct predictions for this part (just the non-overlapping part was taken into account to use every part of the input only "once" for training).

This PR excludes spans that are completely out of the valid window (the not overlapping part). However, we still take predicted spans into account, that are only partly in the valid window because they would get lost otherwise. But this has to be taken into account for further post-processing steps since it may produce the same span multiple times (i.e. from the end of one window and from the beginning of the next one). Any de-duplication is not handled here.