Fix M5 split-K NAX dispatch cliff

[PhilipJohnBasile]

Fixes #4198.

Summary

On M5, the split-K heuristic routes one-M-tile transposed QMMs in the
N=6656..8192 band to the plain Metal qmm_t_splitk kernel. The fallback
qmm path is NAX-capable and is faster for the measured two-way split.

This change:

• factors the existing fallback-QMM NAX eligibility check so the split-K
  decision uses the exact same predicate;
• bypasses split-K only on the measured applegpu_g17s M5 Max class for
  single-batch, transposed affine 4-bit/group-64 QMM with one M tile,
  N=6656..8192, K divisible by 128, and a finalized split count of two;
• leaves all other devices, modes, shapes, and split counts on the existing
  path.

The narrow shape/device gate is intentional. Prior measurements show split-K
still wins in other regions, and this PR only changes the band demonstrated in
#4198. It composes with #4171 if that NAX small-M tiling change lands.

Validation

• Clean editable build from a8e24f202: passed.
• git diff --check, clang-format, Black, Python syntax: passed.
• CPU test_quantized: 37 passed, 3 skipped.
• Pure C++ dispatch-policy test: 18 assertions passed.
• Targeted M5 numerical regression, including M=32/N=6656/K=6656 and adjacent
  N/M boundaries: passed.
• Full M5 Max GPU test_quantized: 37 passed.
• M5 Max ABBA performance sweep: passed; details below.

M5 Max performance

Apple M5 Max (applegpu_g17s), macOS 27.0, bfloat16 activations, affine
4-bit/group-64 weights. Baseline and candidate were separate clean builds from
the same a8e24f202 source. The Metal library hash was identical; only the host
libmlx differed. Each cell used 64 dependent operations, warmup, and seven
samples per arm in baseline/candidate/candidate/baseline order.

Shape	M	baseline	candidate	speedup
K=6656, N=6656	32	251.4 us	196.4 us	1.28x
K=6656, N=7168	32	275.1 us	199.8 us	1.38x
K=6656, N=7680	32	277.4 us	202.6 us	1.37x
K=6656, N=8192	32	299.6 us	201.8 us	1.48x
K=4096, N=8192	32	187.7 us	126.1 us	1.49x
K=5120, N=8192	32	232.0 us	158.5 us	1.46x
K=8192, N=8192	32	361.1 us	241.8 us	1.49x
K=6656, N=8224 (control)	32	204.0 us	204.6 us	1.00x

Across all 33 affected M<=32 cells, speedup was 1.22–1.49x (median
1.37x). Across 24 M>=33 or N=8224 controls, the ratio was 0.992–1.012x
(median 0.999x).

The M=33/M=32 ratio at N=6656/7168/7680/8192 moved from
0.781/0.733/0.732/0.682 on the baseline to 1.005/1.002/1.000/1.007 on the
candidate, eliminating the dispatch cliff. Baseline and candidate ABBA arm
drift stayed below 1.4% and 1.9%, respectively.

[zcbenz]

Thanks for the PR but currently we are only focusing on real cases happened during inference otherwise we wouldn't be able to review all the improvements, so I'm closing this PR.