trunk/3e4e742a98ad43267eb5c5d128b53a9df9386d13: [Pipeline] Release forward outputs after sends complete (#196963)

发生了什么

Store a GradientEdge for each releasable, grad-requiring pipeline output so backward does not require the original tensor object. Track one pending batched send per output slot. When an existing schedule wait completes that send, release the send operations and the stage's output reference. Keep unsupported output types, DDP, custom split backward, and last-stage outputs on their tensor-backed ownership paths. Forwa…

摘要按规则整理自下方来源原文

来源