fix: remove memory_efficient_threshold incompatible with MLX 0.22+

[PatrickHeneise]

Summary

• Removes memory_efficient_threshold from the two scaled_dot_product_attention call sites in mmdit.py — this parameter was removed in MLX 0.22+ and causes a TypeError at the start of denoising
• Guards DiffusionKitInferenceContext.os_spec() in __init__.py with a try/except — it raises IndexError on some macOS configurations where the system profiler output doesn't match the expected tab-delimited layout, crashing generation after the image is already decoded (the image was never saved)

Repro

TypeError: scaled_dot_product_attention(): incompatible function arguments
...
Invoked with types: kwargs = { q: mlx.core.array, k: mlx.core.array, v: mlx.core.array, scale: float64, memory_efficient_threshold: int }

and after patching mmdit.py:

IndexError: list index out of range
  File ".../argmaxtools/test_utils.py", line 596, in os_spec
    line.rsplit("\t\t")[1]

Test plan

• Run diffusionkit-cli with argmaxinc/mlx-stable-diffusion-3.5-large on MLX 0.31.1 — image generates and saves successfully
• Verify no regression on older MLX versions (flash attention path no longer explicitly gated, but MLX handles memory management internally)

🤖 Generated with Claude Code