[CUDA] Avoid copying transposed inputs in row reductions

[Magnushst]

Proposed changes

CUDA reductions currently materialise every non-contiguous input as a contiguous temporary, even when GeneralContiguousReduce can compute the non-reduction-axis offsets directly.

This change uses the direct row-reduction path for GeneralContiguousReduce inputs when the output contains at least 1,024 elements. General reductions, broadcasted inputs, and smaller outputs retain the existing copy path. This removes one full-input allocation, copy, and kernel launch for sufficiently parallel transposed reductions.

A C++ test checks exact CPU/GPU equivalence for the transposed layout handled by the new path.

Validation

• CUDA C++ suite: 262 test cases and 3,412 assertions passed.
• Focused CUDA reduction test: 1 test case and 11 assertions passed.
• Python reduction suite: 10 tests passed.
• pre-commit run --all-files passed.
• git diff --check passed.

The broader Python CUDA suite did not complete because CUDA 13 runtime JIT compilation could not locate cute/tensor.hpp. This header-discovery problem is tracked by #3995; the dedicated reduction suite and complete C++ CUDA suite passed.

Apple Silicon wasn't tested here.

Performance

Environment:

• NVIDIA GeForce RTX 4070 Laptop GPU, compute capability 8.9
• Debian under WSL2, kernel 5.15.167.4
• CUDA 13.2.86
• GCC 14.2.0
• CMake 3.31.6
• Release build
• Baseline 8d666298652fdac2e7727ecdcf507b1d199bba16

The benchmark interleaved the previous copy-plus-reduce path with the direct path in one warmed process. It used 1,000 warm-up evaluations per path, followed by 15 alternating rounds of 500 evaluations. Reported values are medians of the per-round means.

Transposed input	Input size	Copy + reduce	Direct reduce	Change
16x4096x64, axis 1	16 MiB	0.171558 ms	0.038077 ms	77.81% faster
32x2048x128, axis 1	32 MiB	0.424505 ms	0.048314 ms	88.62% faster
16x8192x128, axis 1	64 MiB	0.860658 ms	0.307498 ms	64.27% faster
32x8192x128, axis 1	128 MiB	1.765737 ms	0.611335 ms	65.38% faster

No eligible measured case regressed. Outputs below the 1,024-element threshold retain the existing path. No API or documentation changes are required.

Checklist

Put an x in the boxes that apply.

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)