Beyond Confidence: Reliable Models Should Also Consider Atypicality