Use matrix kernels for global-scale gather_qqmm

[dhiltgen]

This wires up the new gather_qmm kernels to optimize gather_qqmm with global scale.

Performance

Using the NVIDIA Model Optimizer on Qwen/Qwen3.6-35B-A3B with nvfp4_mlp_only p2048/g128

	main prompt tps	this PR prompt tps	main generation tps	this PR generation tps
M5 Max	1,257.0	3,105.3	81.9	81.7
M3 Ultra	1,627.3	2,673.2	68.5	68.6

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: co-developed with coding agent

[zcbenz]

The tests are failing for CUDA backend, not sure of the reason, we can probably disable the test for CUDA for now and create an issue for it so we can look into it later.

[zcbenz]

It seems that the CI failure is just a numeric precision issue, I think we can just allow for a bigger error in CUDA.

[nastya236]

I think it is happening, because we don't have proper global scaling for cuda.. In metal gather_qqmm always uses gather_qmm (because for metal we don't have quantized x quantized instructions). Because we plan to switch to grouped_mm instead of gather_mm (and grouped_qqmm instead of gather_qqmm) specifically on cuda, I think for now we can just skip this test on cuda and fix this with proper scaling in the future grouped_qqmm. What do you think?

[zcbenz]

Sounds good to me!