Improve pre-M5 metal memory usage for SDPA D512

[dhiltgen]

This helps improve the memory usage of D512 attention on M4 and older GPUs with models such as gemma4.

Testing with mlx-community/gemma-4-31b-it-nvfp4 and mlx-community/gemma-4-12B-it-nvfp4 via mlx_lm.benchmark g128

System	Model	Prompt	Build	Prompt tps	Generation tps	Peak memory
M4 Pro, 64 GB	Gemma 4 12B	16K	main	246.530	29.410	9.736 GB
M4 Pro, 64 GB	Gemma 4 12B	16K	this branch	246.767	29.418	9.032 GB
M3 Ultra, 512 GB	Gemma 4 31B	16K	main	326.389	26.949	23.952 GB
M3 Ultra, 512 GB	Gemma 4 31B	16K	this branch	327.970	26.930	22.531 GB
M3 Ultra, 512 GB	Gemma 4 12B	65K	main	664.250	53.256	13.919 GB
M3 Ultra, 512 GB	Gemma 4 12B	65K	this branch	710.878	53.681	10.290 GB

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