Load global scales in qmm_t kernels

[dhiltgen]

Compliments #4481 by adding global scale to the qmm_t kernels.

Performance

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

GPU	upstream/main prompt tps	this PR prompt tps	this + #4481
M5 Max	1,267.5	1,365.9	4,256.2
M3 Ultra	1,603.2	1,683.5	2,819.9

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

[dhiltgen]

A quick helper to load the model in mlx-lm
modelopt_qwen36_mlx_lm.py

[nastya236]

Very nice! I am a very surprised it gives such a big improvement for prefill.. So before scaling was expressed as multiplication of the output after matmul I guess. Do you see any improvements on decode? Probably not, right?
Do you have hf nvfp4 model in mind already quantized that I can use to test the speed up?

[nastya236]

Nit:
I think you can simplify the big if clause with:

  kernel = get_qmm_nax_kernel_wrapped(
        ...
        global_scale.has_value());

[dhiltgen]

To clarify, both this plus #4481 are where the larger win happens, and yes, it's really just prefill. Decode is effectively unchanged. I could be mistaken, but I think the CI failure on the other PR is exposing existing platform-specific CUDA numerical variance, so not sure how you'd like me to address that.