A density estimation perspective on learning from pairwise human preferences