Revealing the Proximate Long-Tail Distribution in Compositional Zero-Shot Learning