Debunking Optimization Myths in Federated Learning for Medical Image Classification