Run the backward-only all-reduces only when a gradient can flow
One helper, _all_reduce_backward, applies _AllReduceBackward only when grad is
enabled and the input requires it, like the TP norm hooks (#49975). The MoE TP
experts, the MLA rope projection and the EP inputs no longer run it at inference
(~1.8us of host time per call in eager decode); the sites that already checked
use the helper too.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>