Exploring Safety-Utility Trade-Offs in Personalized Language Models