B=1 response-only length-normalized mean log-prob from pre-computed final-hidden h [1, T, hidden].
Mirror of responseOnlyCe but returns mean response logp (scalar wrapped in [1]) instead of
CE loss. Used by SegmentedBackwardOrpo's head VJP. Disposes nothing; caller owns h and result.
B=1 response-only length-normalized mean log-prob from pre-computed final-hidden h [1, T, hidden]. Mirror of responseOnlyCe but returns mean response logp (scalar wrapped in [1]) instead of CE loss. Used by SegmentedBackwardOrpo's head VJP. Disposes nothing; caller owns h and result.