Towards Human-AI Complementarity with Prediction Sets