Trainability barriers and opportunities in quantum generative modeling