Bump mlx-c to v0.6.0 (MLX 0.31.1)

[dusterbloom]

Summary

• Bump the mlx-sys/src/mlx-c submodule from v0.5.0 (MLX 0.30.6) to v0.6.0 (MLX 0.31.1)
• Adapt quantize/dequantize to the new C signatures: MLX 0.31 adds an optional global_scale array parameter to both (used by the new two-level-scaled MXFP4 path). A fresh mlx_array_new() handle is passed, which the C API treats as null — matching the existing convention for optional arrays (e.g. biases in quantized_matmul).

Motivation

MLX 0.30.6's scheduler deadlocks on large GDN-hybrid models on smaller unified-memory boxes (repro: Qwen3.8-27B 4-bit, 15.2 GiB checkpoint, M4 32 GB — first wide prefill wedges eval_impl on a condition variable forever; an io_connect_method GPU submission never completes). 0.31.1 fixes this class. Details and a sample backtrace: measured while porting the model to the higgs inference server.

Testing

• cargo build clean for mlx-sys and mlx-rs
• Full model load + serve + MTP speculative decode verified on M4 (macOS 26), including the quantize/dequantize paths exercised by checkpoint loading
• No other API surface changed between 0.30.6 → 0.31.1 in the C bindings used by mlx-rs

Follows the pattern of #98 (v0.4.1 → v0.5.0 bump).