Subtleties in the trainability of quantum machine learning models