Improve metal memory usage for SDPA D512

[dhiltgen]

This helps improve the memory usage of D512 attention on NAX GPUs with models such as gemma4

Testing with mlx-community/gemma-4-31b-it-nvfp4 and mlx-community/gemma-4-12B-it-nvfp4

Model	Prompt	main peak	this branch peak	Memory saved
Gemma 4 31B	16K	23.9515 GB	22.0368 GB	1.9147 GB (7.99%)
Gemma 4 12B	16K	9.7047 GB	8.7595 GB	0.9452 GB (9.74%)
Gemma 4 12B	65K	13.8004 GB	9.5775 GB	4.2228 GB (30.60%)

Both prompt and generation performance are unchanged.

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