[core][distributed] fix custom allreduce in pytorch 2.5

[core][distributed] fix custom allreduce in pytorch 2.5 #1693