[CUDA] Add global scale support to gather_qmm

[dhiltgen]

This adds support for global scales in gather_qmm on cuda.

mlx-lm perf data with qwen 3.6 35B tested on a 5090. (p2048/g128)

model / system	prompt tps	gen tps
mlx-community/Qwen3.6-35B-A3B-nvfp4 (no global scale)	394.3	75.9
Model Optimizer export (with global scale)	404.8	70.5

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

Attached wrapper script loads the Model Optimizer exports and wires up gather_qmm global scale, and adjusts dtypes for a fair comparison with the standard community model.
modelopt_nvfp4_mlxlm.py

[nastya236]

Thanks! Why "optimizes performance"? Did you notice a speed up?

General overall note: after giving this some more thought, I realised that dequantization during inference would be correct if and only if it matches the training. I believe there are two possible ways to produce nvfp4 weights:

1. quantize(W) -> use nvfp4 x nvfp4 instructions for matmul and fold global scale into epilogue — in this case we also must fold scale into epilogue during inference as well (like in this PR)
2. If the global scale was applied when dequantizing the weights, before casting to BF16 and performing the matrix multiplication in bf16 (quantization aware training on sm < 100), we should follow that same sequence during inference — this is done in [Metal] global scale for qmm #4458

Given that in the future all quantized operations during training will be native, I think it is okay to agree on folding it into epilogue. However, probably we should do the same for metal..

[nastya236]

Some tests are failing can you have a look?

[dhiltgen]

Why "optimizes performance"? Did you notice a speed up?

On further testing, I couldn't reproduce any speedup in the second commit so I've dropped it and the claim.