Policy-based Tuning of Autoregressive Image Models with Instance- and Distribution-Level Rewards