CompA: Addressing the Gap in Compositional Reasoning in Audio-Language Models