Quantifying Interpretability in CLIP Models with Concept Consistency