Pin mlx dependency to PrismML 1-bit affine quantization fork

Point at PrismML-Eng/mlx@prism which adds 1-bit affine
quantization Metal kernels, enabling Bonsai-8B-mlx-1bit
(1.2 GB Qwen3-8B) to run locally on Apple Silicon.

Temporary pin until ml-explore/mlx#3161 merges upstream.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: lyonsno

setup.py

@@ -24,7 +24,7 @@
     url="https://github.com/ml-explore/mlx-lm",
     license="MIT",
     install_requires=[
-        f"mlx>={MIN_MLX_VERSION}; platform_system == 'Darwin'",
+        "mlx @ git+https://github.com/PrismML-Eng/mlx.git@prism ; platform_system == 'Darwin'",
         "numpy",
         "transformers>=5.0.0",
         "sentencepiece",