[Distributed] Use custom overlapping method for backward chunks

[lshpku]

PR Category

Auto Parallel

PR Types

Improvements

Description

在 DualPipeV 中，当执行 backward_pass 时，不直接调用 backwad_chunk.backward，而是调用模型的 overlapped_forward_backward 方法，但只传入 backward_chunk，不传入 forward_chunk：

model.overlapped_forward_backward(
    forward_chunk=[],
    backward_chunk=backward_chunk
)

这样允许模型自主编排 backward_chunk，实现 backward 之间的自 Overlap

该 PR 为 兼容性升级，不会破坏现有模型的兼容性，模型若未实现 backward 的自 Overlap 机制，则会跑原来不 overlap 的 backward，性能不变

DSV3 实现自 Overlap 机制：PaddlePaddle/PaddleNLP#11003

Pcard-85711

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请关注后续CI自动化测试结果，详情请参考Paddle-CI手册。
Your PR has been submitted. Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[codecov-commenter]

Codecov Report

❌ Patch coverage is 33.33333% with 2 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@b95cb68). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...addle/distributed/fleet/meta_parallel/dualpipev.py	33.33%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop   #74891   +/-   ##
==========================================
  Coverage           ?   33.33%           
==========================================
  Files              ?        1           
  Lines              ?        3           
  Branches           ?        0           
==========================================
  Hits               ?        1           
  Misses             ?        2           
  Partials           ?        0           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.