[WIP][Dy2St] Move scope cache to cpp side

[SigureMo]

PR Category

Execute Infrastructure

PR Types

Improvements

Description

同 #71505，为逻辑优化，将 scope cache 逻辑移到 C++，以和执行器使用相同的 cache 计算逻辑

PR Overview

This PR refactors the dy2static partial program implementation by moving the scope cache logic to the C++ side, while also cleaning up unused CUDA graph configuration.

• Replaces _legacy_C_ops.pir_run_program with _C_ops.run_program.
• Removes CUDA graph vector initialization and related attribute settings.
• Renames the backward function from _append_backward_desc to _append_backward.

Reviewed Changes

File	Description
python/paddle/jit/dy2static/pir_partial_program.py	Scope cache logic moved to C++ side; cleanup of CUDA graph-related code and renaming of backward function.

Copilot reviewed 6 out of 6 changed files in this pull request and generated no comments.

Comments suppressed due to low confidence (2)

python/paddle/jit/dy2static/pir_partial_program.py:717

• Confirm that the updated _C_ops.run_program API no longer requires the CUDA graph vector argument. Its removal here could lead to unexpected behavior if the C++ implementation still expects this parameter.

_C_ops.run_program(

python/paddle/jit/dy2static/pir_partial_program.py:1115

• Verify that removing the CUDA graph capture mode attributes does not inadvertently remove required configurations for environments that rely on CUDA graph functionalities.

if self._cuda_graph_capture_mode:

[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.