Airway Labeling Meets Clinical Applications: Reflecting Topology Consistency and Outliers via Learnable Attentions