[Eager] fix recompute for stop_gradient and inpalce

[sljlp]

PR types

Bug fixes

PR changes

Others

Describe

Fix recompute for stop_gradient and inpalce.

In some cases a subset of outputs is identity to the subset of inputs, which is inplace operating. When the inputs' stop_gradient is True, an error will occurs because the stop_gradient=True and inpalce-op are not supported at the same time. The solution is to mark the inputs non_differentiable if its stop_gradient is True.

• Note: if not marked non_differentiable, all output tensors'
  attribute stop gradient will be reset to False in c++ backend . code:

  Paddle/paddle/fluid/pybind/eager_py_layer.cc

  Line 388 in 9d62efb

  outputs_autograd_meta[i][j]->SetStopGradient(false);

[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.

[LiYuRio]

LGTM