How to Learn when Data Reacts to Your Model: Performative Gradient Descent