On Class Separability Pitfalls In Audio-Text Contrastive Zero-Shot Learning