feat!: align max unpool with PyTorch Python API

[voltjia]

Summary

• Align max_unpool2d and max_unpool3d with PyTorch's Python geometry order using one full public interface per operator.
• Add shared output-size resolution, a handwritten max_unpool2d Torch adapter, overload-aware vector code generation, and focused CPU/NVIDIA coverage.

Motivation

The existing operators expose ATen's resolved output_size, stride, padding contract rather than the public functional API. Public callers therefore cannot use PyTorch's geometry contract directly. The generated max_unpool2d.out adapter also cannot implement Python-level geometry resolution by itself.

No standalone issue is associated with this alignment follow-up.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

Host: ssh nvidia
Image: accelerator-dev/nvidia:latest
PyTorch: 2.10.0a0+b4e4ee8

cmake ... -DWITH_CPU=ON -DWITH_NVIDIA=ON -DWITH_TORCH=ON \
  -DINFINI_OPS_OPS=max_unpool2d,max_unpool3d \
  -DINFINI_OPS_TORCH_OPS=max_unpool2d,max_unpool3d
cmake --build build-nvidia --target ops -j 16
Build passed; codegen emitted 1 max_unpool3d overload and the handwritten max_unpool2d adapter was linked.

python -m pytest tests/test_generate_wrappers.py -q
14 passed

python -m pytest tests/test_torch_ops.py -k max_unpool --devices cpu nvidia -q
36 passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Yes	targeted build passed	2D/3D geometry cases passed
NVIDIA	Yes	targeted build passed	2D/3D geometry cases passed
Iluvatar	No	N/A - not affected	N/A - not affected
MetaX	No	N/A - not affected	N/A - not affected
Cambricon	No	N/A - not affected	N/A - not affected
Moore	No	N/A - not affected	N/A - not affected
Ascend	No	N/A - not affected	N/A - not affected

Benchmark / Performance Impact

N/A. Geometry is resolved once per operator construction/call; the underlying ATen kernels are unchanged.

Notes for Reviewers

API alignment

Interface	Signature	Alignment basis
InfiniOps 2D	max_unpool2d(input, indices, kernel_size, stride, padding, output_size, out)	PyTorch Python geometry order; explicit out remains last per CONTRIBUTING.md.
PyTorch 2D	F.max_unpool2d(input, indices, kernel_size, stride=None, padding=0, output_size=None)	Pinned PyTorch source
InfiniOps 3D	max_unpool3d(input, indices, kernel_size, stride, padding, output_size, out)	PyTorch Python geometry order; explicit out remains last per CONTRIBUTING.md.
PyTorch 3D	F.max_unpool3d(input, indices, kernel_size, stride=None, padding=0, output_size=None)	Pinned PyTorch source
Deprecated 3D compatibility	max_unpool3d(input, indices, output_size, stride, padding, out)	Preserves the pre-alignment 3D interface for migration.

• No prefix/default convenience overloads are added; C++ callers supply the optional values explicitly.
• The old 2D and new 2D signatures have the same C++ type sequence, so they cannot coexist as overloads. This is the direct breaking replacement reflected by feat!.
• max_unpool2d is removed from scripts/torch_ops.yaml because its ATen schema exposes resolved output_size, not the public Python geometry contract. The handwritten adapter resolves geometry and then calls at::max_unpool2d_out.
• Ruff 0.15.22 and clang-format 21.1.8 checks passed.