DPO loss for one batch given precomputed reference log-probs ([B] each).
Runs the two policy forwards (adapter active) and reduces to the scalar
mean over the batch of -logσ(adv) (matching dpo.py _policy_dpo_loss).
Returns a scalar MlxArray (caller owns).
DPO loss for one batch given precomputed reference log-probs ([B] each). Runs the two policy forwards (adapter active) and reduces to the scalar mean over the batch of -logσ(adv) (matching dpo.py _policy_dpo_loss). Returns a scalar MlxArray (caller owns).