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