Too Good to be Bad: On the Failure of LLMs to Role-Play Villains