On the Limitations and Prospects of Machine Unlearning for Generative AI