Add global scale for nvfp4 quantized MoEs

[nastya236]

Follow up of ml-explore/mlx#4458 that adds global scale to gather_qmm with nvfp4 weights.
[should be merged after mlx pr]

Tested with mlx-community/Qwen3.6-35B-A3B-nvfp4:

mlx_lm.chat --model mlx-community/Qwen3.6-35B-A3B-nvfp4 --max-tokens 1024

Input: What is the highest mountain? Be brief.
Output:
Here's a thinking process:

1. Analyze User Input:

• Question: "What is the highest mountain?"
• Constraints: "Be brief."
• Key requirement: Direct, concise answer without elaboration or reasoning.

2. Identify Core Fact:

• The highest mountain above sea level is Mount Everest.
• Height: 8,848.86 meters (29,031.7 feet) above sea level.

3. Apply Constraints:

• "Be brief." → Keep it to one short sentence or just the name.

4. Formulate Response:

• "Mount Everest." (or "Mount Everest, 8,848.86 meters above sea level.")

5. Final Check:

• Matches fact? Yes.
• Brief? Yes.
• No extra thinking/reasoning? Yes.
• Ready. Output: "Mount Everest."✅

Mount Everest.

[gabe-l-hart]

This branch was proposed as a fix for the underflow issues I'm seeing with Granite in #1836 (both dense and moe). I tested it out, but it doesn't seem to fix the issue for either model, and the MoE path raises a stack trace: #1836 (comment)