Retune MLX transposed quant matmul on Metal

[ry2009]

What changed

• keep QmmTransposed64x64 on the zero-point path only
• use an MLX-only WM=4, WN=1 split in qmm_impl
• retile QmmTransposed from 32x32 to 160x32

Why

Llama-3.2-3B-Instruct-4bit prefill on Apple was still spending most of its time in transposed quant matmul. The MLX 4-bit path was taking a bad kernel family and a weak transposed tile shape for this workload.

This patch keeps the change set narrow and only moves the MLX path onto the kernel shapes that benchmarked well on Apple.

Impact

Matched Apple blocker benchmark: Llama-3.2-3B-Instruct-4bit, 10049 prompt tokens, 32 generated tokens, tiered_q4:256.

Variant	TTFT (s)	Prompt tok/s	Gen tok/s	Memory (GB)
baseline clean tiered_q4	47.33	212.34	23.33	2.68
disable 64x64 family	47.24	212.73	23.60	2.62
add MLX-only WM=4, WN=1	47.16	213.11	24.28	2.58
final 160x32 transposed QMM	38.99	257.73	23.95	2.48

Negative ablations on the same Apple box:

Variant	TTFT (s)	Prompt tok/s	Gen tok/s	Memory (GB)
160x32 with old 2x2 split	40.53	247.96	23.36	2.58
192x32	43.34	231.91	23.64	2.50
256x32	43.71	229.94	24.01	2.68

So the main win comes from the transposed 160x32 tile, with the MLX-only 4x1 split adding a smaller but real gain.

Validation

• PATH="$HOME/.rustup/toolchains/1.93.0-x86_64-apple-darwin/bin:$PATH" cargo check -p uzu --no-default-features --lib
• Apple: cargo build -p cli --release
• Apple benchmark result files already present on the remote box:
  • llama3b_4bit_tiered256_x12_long32_clean.results.json
  • llama3b_4bit_wm4_r2.results.json
  • llama3b_mlx_qmm_transposed_160x32_r2.json
  • llama3b_mlx_qmm_transposed_160x32_2x2.json
  • llama3b_mlx_qmm_transposed_192x32.json
  • llama3b_mlx_qmm_transposed_256x32.json

[ry2009]

Superseded by #323, which is the same matmul change rebased onto current main for clean benchmarking.

[ry2009]

Closing in favor of #323.