Split a DpoBatch row (B=1) into the shared prompt prefix + the two response
continuations. The chosen/rejected sequences must share an identical prompt
(the masked prefix where mask==0); response tokens are mask==1. Returns null
if the prompts differ or either response is empty (caller falls back to the
two-forward orpoLoss).
Split a DpoBatch row (B=1) into the shared prompt prefix + the two response continuations. The chosen/rejected sequences must share an identical prompt (the masked prefix where mask==0);
responsetokens are mask==1. Returns null if the prompts differ or either response is empty (caller falls back to the two-forward orpoLoss).