Harder or Different? Understanding Generalization of Audio Deepfake Detection