ggml webgpu: Fix bug in dispatching large matrix-vector multiplication

[reeselevine]

Bug fix for calculating overflowing workgroup sizes for large matrix-vector multiplication batches. Should fix failures from new tests in #19519.

This approach isn't ideal because it may over-provision workgroups by quite a bit, a better strategy is the one proposed for Vulkan in #19509, but this will work for now.

[reeselevine]

@jeffbolznv can you approve this PR or should I ask someone else to?

[jeffbolznv]

I don't quite follow how this is fixing it. But if you explain it, I can try to review.

[reeselevine]

Sure, this was actually just a bug in the logic to handle cases where the number of workgroups that should be launched exceeds the X dimension limit. If it does overflow, we should launch the maximum number of workgroups in the X dimension, and then in the Y dimension, the total necessary workgroups divided by the maximum number of workgroups in the X dimension (total_wg/maxComputeWorkgroupsPerDimension). This way we have enough total workgroups launched, with the workgroup ids linearized in the shader.

This does overprovision the number of workgroups, for example if there are 65536 workgroups needed, we will end up launching 2 * 65335, but extra workgroups exit immediately. The TODO comment is to improve this in the future.