Explainability in Generative Medical Diffusion Models: A Faithfulness-Based Analysis on MRI Synthesis