On CUDA, take every rank's late-layer MoE gather size from the scheduled
per-rank trim counts, which ride the existing scheduler all-gather, and switch
and restore them with #43077's late-layer DP helpers; ROCm keeps its eager
count gather. Make the helpers' count tensor copy asynchronous so the switch
does not synchronize the stream. Restrict to attention TP 1 and TP MoE.
(cherry picked from commit a6c2d5f40efec4b491fcdd2dbcedf14ec70d3861)