A Prompt Array Keeps the Bias Away: Debiasing Vision-Language Models with Adversarial Learning