Controlling What You Share: Assessing Language Model Adherence to Privacy Preferences