[Fleety_12][Precision Depth Alignment] Fix paddle.nn.functional.affine_grid forward and backward gradient calculation

[zhengshengning]

PR Category

Operator Mechanism

PR Types

New features

Description

修改内容：

1. 去掉了原来 cudnn 的实现方式（通过跟相关负责人沟通得知，原来的实现是临时方案，且无法与torch精度对齐）；
2. 分别对4D和5D手写了对应的cuda kernel实现；

前向传播

1. 生成基础网格：

   • 4D: 为每个(H, W)位置生成归一化坐标(x, y, 1)
   • 5D: 为每个(D, H, W)位置生成归一化坐标(x, y, z, 1)

2. 仿射变换：

   • 4D: [N, H×W, 3] × [N, 3, 2] = [N, H×W, 2]
   • 5D: [N, D×H×W, 4] × [N, 4, 3] = [N, D×H×W, 3]

反向传播

1. 梯度计算：
   • 4D: [N, 3, H×W] × [N, H×W, 2] = [N, 3, 2] → 转置为 [N, 2, 3]
   • 5D: [N, 4, D×H×W] × [N, D×H×W, 3] = [N, 4, 3] → 转置为 [N, 3, 4]

两者遵循相同的计算模式：生成基础网格 → 应用仿射变换 → 输出变换后的网格坐标。

测试结果：

1. 测试集的 paddle.nn.functional.affine_grid 的4D case 全部对齐（10个）；
2. 新增 10 个 5D的测试case全部通过（10个）；

devPR:#76066

[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

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (fleety_12@7320f14). Learn more about missing BASE report.

Additional details and impacted files

@@              Coverage Diff              @@
##             fleety_12    #76128   +/-   ##
=============================================
  Coverage             ?   100.00%           
=============================================
  Files                ?         1           
  Lines                ?         1           
  Branches             ?         0           
=============================================
  Hits                 ?         1           
  Misses               ?         0           
  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.