Coarse-to-Fine Q-attention with Learned Path Ranking