webgpu : add CONV_2D_DW (depthwise conv2d) kernel

[m1el]

Implement GGML_OP_CONV_2D_DW for the WebGPU backend, ported from the Vulkan backend's conv2d_dw.comp.

Assisted-by: Claude Opus-4.8

Overview

Addresses #14909

This PR implements GGML_OP_CONV_2D_DW for the WebGPU backend.
One of my projects based on ggml requires depthwise conv2d.
It's possible to work around this by implementing this function using other primitives. Having feature parity between the backends is a better solution.

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES.
  Claude-4.8 was used to port the shader from vulkan to wgsl.
  The work performed is mostly mechanical translation from one language to another.
  I have read the code and understood it. I am ready to talk about the code if need to show my understanding.

[yomaytk]

Thanks! but we should update docs/ops.md when new ops are added, so could you update it? Please see the description in ops.md.

[m1el]

I did attempt updating the docs/ops.md as described. The issue I've had is it is a bit outdated, so the table change is huge.
I'll commit the updated support matrix, tell me later if you want the change.

[m1el]

Regenerated docs/ops/*, but the diff is now quite bloated.
I updated as many backends as I could because new ops were added and because I have access to those backends.

[yomaytk]

Ideally, each CSV should completely reflect the actual status of the ops implementation across all backends. But for this PR, I think updating only WebGPU.csv is enough — could you only update that file?

[m1el]

Only updating WebGPU.csv, regenerated docs, done.

[yomaytk]

Some CI tests failed, but they don't seem to be related to this change.

@reeselevine, thanks for working on #25943 and #25931. These PRs seem to address those failures — is that right?

[reeselevine]

@yomaytk yep I'm trying to get the tests passing again in those PRs :)