The MedPerturb Dataset: What Non-Content Perturbations Reveal About Human and Clinical LLM Decision Making