Better than classical? The subtle art of benchmarking quantum machine learning models