Improve metal memory usage for SDPA D256

[dhiltgen]

This helps improve the memory usage of D256 attention on M4 and older GPUs with models such as qwen, as well as a small prompt speedup.

Testing with mlx-community/Qwen3.8-27B-nvfp4 and mlx-community/Qwen3.5-2B-nvfp4 via mlx_lm.benchmark

System	Model	Prompt length	Build	Prompt tps	Generation tps	Peak memory
M4 Pro, 64 GB	Qwen3.8-27B	16K	main	115.8	14.3	46.92 GB
M4 Pro, 64 GB	Qwen3.8-27B	16K	this branch	125.5	14.3	35.46 GB
M1, 16 GB	Qwen3.5-2B	4K	main	444.0	47.1	2.78 GB
M1, 16 GB	Qwen3.5-2B	4K	this branch	454.6	47.0	2.70 GB

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