Reorder output keys: entities before json_structures (atomic NER layer first, matching bc5cdr)

Rename entity-type keys to human-readable names (PER->person, GPE->geopolitical entity, ...)

Add M2E2 (text) EE+NER dataset — text_only subset, 6013 test records

initial commit