Run 2D MaxPool on the Metal GPU

[czoli1976]

MaxPool had no Metal kernel, so it round-tripped its tensor through the host in the middle of a GPU graph. This adds a kernel that pools two spatial axes with every axis stride passed explicitly, so NCHW and NHWC share one path.

Stacked on #2554, #2555 and #2556 — this branch carries their commits and this PR's own change is the last one. It only depends on #2556 for the benchmark story, not for the code, so it can also be taken on its own if that's easier.

Effect

The point of this one is not the kernel's own throughput, it's what it removes. u2netp (rembg), 1×3×320×320, --metal -O, min of 3 interleaved rounds:

	host round-trips (DeviceSync*)	end-to-end
main	150	560.7 ms
+ Metal Resize (#2556)	74	460.7 ms
+ Metal MaxPool (this)	8	359.8 ms

The graph is now essentially GPU-resident, and that is what makes Metal win on this model: 359.8 ms vs 419.1 ms on CPU. It lost before either kernel (560.7) and still lost with Resize alone (460.7) — the residual host transfers dominated. Resize was necessary, MaxPool was what flipped it. Together, 1.56× on the Metal path.

Outputs match the CPU runtime to 5.96e-7 max abs across all seven model outputs.

Notes

Padding positions are skipped rather than compared, so a window lying entirely in the padding keeps the CPU op's min_value() result.

The translator registers for both MaxPool and OptMaxPool. The Metal transform runs before the optimize pass, so the node is still MaxPool at translation time even though a -O dump shows OptMaxPool — registering only for the optimized op silently never fires. OptMaxPool is added to core::ops::cnn's public re-exports for this.

Two limits worth knowing:

• Scoped to rank 4, two spatial axes, no index outputs, concrete shapes. 1D/3D pooling, with_index_outputs and symbolic shapes aren't translated and stay on CPU.
• One thread per output element with a 1×1×1 threadgroup, matching the existing Gather/DiagGather kernels. A tuned threadgroup size probably has more in it; I left that alone because the transfer elimination is the dominant term here and I didn't want to tune on a single model.

Testing

test-metal 22762 passed / 0 failed (558 of them pooling node tests; I checked the nodes actually land on MetalMaxPool rather than silently falling back to CPU). onnx-tests.sh 1_19_1 and test-unit-core green. cargo fmt --all and cargo clippy clean.

🍍

[czoli1976]

Marking this draft: it overlaps #2553, which I opened earlier and had not checked before writing this. #2553 already adds Metal max and sum pooling, with a better-coalesced NHWC dispatch.

The two are not identical, though — #2553 deliberately leaves NCHW on the CPU path, and this one covers NCHW, which is what an ONNX/PyTorch export like u2netp uses. I will rework the useful part of this as NCHW support on top of #2553 rather than ask anyone to review two pooling kernels.

[czoli1976]

Superseded by #2553, which now carries channels-first max pooling as a second commit. Closing this rather than asking anyone to review two pooling kernels.