Evaluating the Stability of Semantic Concept Representations in CNNs for Robust Explainability