Discovering Generalizable Spatial Goal Representations via Graph-based Active Reward Learning