MedSAE: Dissecting MedCLIP Representations with Sparse Autoencoders