Length-normalized mean log-prob over gathered (possibly non-contiguous) hidden
positions via the FUSED/FLASH head ([M,V]-free analytic backward) when chunk
requests it, else the whole-vocab gatheredLogpMean. Gathers the branch's
response hiddens [M, hidden] from h [1,T,hidden] and hands them to
fusedRespLogpMean — so prefix-sharing composes with the steel flash-CCE head
(one model forward + one head fwd/bwd per branch, all [M,V]-free).
Length-normalized mean log-prob over gathered (possibly non-contiguous) hidden positions via the FUSED/FLASH head ([M,V]-free analytic backward) when
chunkrequests it, else the whole-vocabgatheredLogpMean. Gathers the branch's response hiddens [M, hidden] fromh [1,T,hidden]and hands them tofusedRespLogpMean— so prefix-sharing composes with the steel flash-CCE head (one model forward + one head fwd/bwd per branch, all [M,V]-free).