Bradley-Terry Reward Model Loss

Compute Bradley-Terry pairwise preference loss for chosen and rejected reward-model scores in RLHF.