[Optimization] GridSample is slow and is making InternImage run at less than 1 fps

DOR (Definition of Ready)

GridSample's ONNX-parser decomposition (src/onnx/parse_gridsample.cpp) is a performance bottleneck: for mode="linear", concat+gathernd index materialization measured at 78 ms. A gridsample operator + GPU JIT kernel was developed in a draft PR and lands this gap for mode="linear" in this draft pr. Precedent and file layout are established; remaining work is extending the same operator to mode="nearest" and mode="cubic", all three padding_mode values, and both align_corners settings. Volumetric (5D) GridSample is explicitly rejected by the parser (gridsample_volumetric_*_test) and is out of scope.

Description

As a MIGraphX user compiling models with instance segmentation heads (e.g. Mask R-CNN mask paste), I want GridSample to run as a single fused GPU kernel for all supported modes (nearest, linear, cubic × zeros/border/reflection × align_corners), not just linear, so that no GridSample usage still falls back to the index-materializing decomposition in parse_gridsample.cpp

DOD (Definition of Done)

• gridsample operator supports mode ∈ {nearest, linear, cubic} (already: linear; add: nearest, cubic).
• GPU JIT kernel handles all three modes; nearest has no interpolation (single tap), cubic computes 16 taps (4×4) inline — no gathernd/concat for any of them.
• Perf comparison (old vs. new) reported for at least one shape per mode.
• Correctness verified against torch.nn.functional.grid_sample for nearest and bicubic, same method as the linear PR.
• CHANGELOG.md entry added.

Start by reviewing the draft PR and src/onnx/parse_gridsample.cpp, then inspect the gridsample_volumetric_*_test cases to confirm the 5D exclusion. Extend the established approach for the remaining modes, padding modes, and align_corners settings, compare performance by mode, verify against torch.nn.functional.grid_sample, and add a CHANGELOG.md entry.