Prevent overlapping annotations from LLM evaluators

[kasnerz]

We have no way to prevent LLMs from producing annotations that overlap.

Since we decided not to support that (#31), we should modify the matching algorithm to disallow overlapping spans. Possible solutions include:

• trimming the suffix of the preceding annotation,
• trimming the prefix of the follow-up annotation,
• skipping the overlapping annotation.

[kasnerz]

Update: In general, our matching algorithm ensures that no annotations overlap.

The only thing that is required is preventing matching from scratch. The purpose of this extra tweak was to "save" incorrectly ordered annotations which would otherwise be lost.

But since the feature is undocumented, it makes sense to remove it. It is already removed in the upcoming feature pack:

factgenie/factgenie/models.py

Line 139 in cb834b9

# if current_pos != 0 and start_pos == -1: