Clean overlapping annotation spans and fix evaluation

[monradach]

• Original MedDec annotations contain many overlapped and nested spans (305 files with overlaps, 2,332 overlapping pairs total) 98994_178949_30340_original.pdf
• clean_data.py now sorts annotations by start_offset before saving
• Add cleanup logic for same_class_overlap, same_class_nested, and multi_class_nested spans (remaining overlapping pairs after cleanup: 385) 98994_178949_30340_updated.pdf
• For remaining overlaps, evaluate.py enforces a single label per token like the original script: if multiple spans cover the same token, the first span encountered takes priority
• Fix evaluate.py: improve file ID matching in splits and fix '_' handling in load_gold_annotations