Test-Time Distribution Normalization for Contrastively Learned Vision-language Models