CBVLM: Training-free Explainable Concept-based Large Vision Language Models for Medical Image Classification