[Backend] Fix explicit pooling pads and CUDA TopKV2 largest flag

[Fatemanx]

Summary

• Honor CAFFE explicit pads() for non-global pooling across CPU, CUDA, and OpenCL paths.
• Keep global pooling padding consistent with shape inference by ignoring explicit pads for global pooling.
• Respect CUDA TopKV2 largest so largest=false returns the smallest values.

Related Issue

Related to #4613

Test Plan

• cmake --build build_cuda -j$(nproc) --target run_test.out
• ./build_cuda/run_test.out op/CUDAMaxPoolExplicitPads 0 0 0
• ./build_cuda/run_test.out op/CUDAMaxPoolExplicitPads 2 0 0
• ./build_cuda/run_test.out op/CUDATopKV2LargestFlag 0 0 0
• ./build_cuda/run_test.out op/CUDATopKV2LargestFlag 2 0 0
• ./build_cuda/run_test.out op/TopKV2 0 0 0
• ./build_cuda/run_test.out op/TopKV2 2 0 0
• ./test.sh static

OpenCL execution was not run locally because this environment does not provide an OpenCL executor.

[Qxinyu]

Thanks for the fix!

Verified on x86 Linux + CUDA (RTX 3090): both new tests fail on current master, confirming the bugs this PR addresses.
Root causes confirmed:

1. Pooling: when ONNX explicit pads is present (padType=CAFFE), backends switched padType to VALID without reading pads() values, so pooling ran with zero padding.
2. CUDA TopKV2: mParams.mDescendFlag = boolDescendFlag ? 1 : -1 tested the param pointer instead of the largest() field, so largest=false was always ignored.

Suggestions on test organization:

• Both tests are actually backend-agnostic (they exercise CPU/OpenCL as well), so the CUDA prefix in file/test names is misleading.
• CUDATopKV2LargestFlagTest.cpp: consider merging the largest/smallest cases into the existing test/op/TopKV2Test.cpp (op/TopKV2), which currently only covers the default largest=true path.
• CUDAMaxPoolExplicitPadsTest.cpp: there is no existing 2D forward pooling test in the repo (only PoolGrad/Pool3D/ROIPooling), so this is effectively the first one — consider renaming it to PoolTest.cpp (e.g. registered as op/MaxPoolExplicitPads) so future 2D pooling cases have a natural home.

Thanks again for the contribution!