Add GPU JIT implementation for gridsample operation

[Imeguras]

Motivation

MIGraphX's ONNX parser lowers bilinear GridSample into a concat+gathernd decomposition that materializes explicit index tensors for every output element rather than computing sample positions inline.

For a typical Mask R-CNN mask-paste (100 instances, 28×28→640×640), this produces ~327 MB of baked index literals and takes 78.0 ms on an RX 7800 XT — 51% in concat, 27% in gathernd.

In practice this makes something like InternImage(the smaller models) run at less than 1 fps even with a rocm/hip and other amd dedicated frameworks, which for my project(being an alternative to deepstream is just horrible).

Technical Details

parse_gridsample.cpp's linear_sampler::sample() lowers bilinear GridSample into a concat+gathernd decomposition. For each output element it builds explicit {x0,y0}/{x1,y0}/{x0,y1}/{x1,y1} corner index tensors as ONNX-parser-time literals, then does four gathernd fetches + weighted add. It werks, but ends up baking a lot of memory for index literals into the compiled program.

Fix: a new gridsample operator with a dedicated GPU JIT kernel, following the existing pattern used by nonmaxsuppression/resize/roialign (src/targets/gpu/jit/.cpp + src/targets/gpu/kernels/include/migraphx/kernels/.hpp):

• src/include/migraphx/op/gridsample.hpp — op definition + CPU reference compute(), used by the ref target and as the correctness baseline.
• src/targets/gpu/kernels/include/migraphx/kernels/gridsample.hpp — the fused device kernel. One thread per output element; unnormalizes its own (h,w) into a source coordinate and computes all four bilinear taps directly — no index tensor, no intermediate buffer. align_corners/padding_mode are template/constexpr params so the padding branch is compiled out per-instance.
• src/targets/gpu/jit/gridsample.cpp — hiprtc JIT compiler wiring, mirrors jit/roialign.cpp.
• src/onnx/parse_gridsample.cpp — emits gridsample for the bilinear case; nearest/bicubic/mixed-dtype inputs still fall through to the existing decomposition unchanged. MIGRAPHX_DISABLE_GRIDSAMPLE_OP=1 forces the old path for A/B comparison(im gonna review this once again im still learning migraphx and how it works).
• src/CMakeLists.txt, src/include/migraphx/operators.hpp — op registration.
• test/onnx/parse/gridsample{,_channel,_512x512}_test.cpp — updated to assert the single-instruction lowering instead of the decomposition.

Changelog Category

• • [*] Optimized: Component performance that has been optimized or improved.

[github-actions]

Thank you for your contribution! Since this is an external pull request, a maintainer must review PR and add the "ok-to-test" label if it is approved for testing.

[Imeguras]

Hello! Im new to the contributing to such a big project.

I noticed that one of the tests was building the gridsample by hand is it intentional? I look into others but didnt get my answer.

[Imeguras]

... Found a bug while testing out the performance increases on InternImage
Debugging it right now then will review my code when im not tired from work. should have it by tomorrow