NExtLong: Toward Effective Long-Context Training without Long Documents