Tune M1 quantized matmul dispatch for six-row large shapes

[DeadByDawn101]

Summary

For applegpu_g13s, keep large (D > 4096, O > 4096) affine quantized matmul inputs on the QMV path through six rows by changing the matrix-dispatch boundary from 6 to 7.

This is intentionally limited to the measured M1 g13s architecture and leaves every other architecture/shape threshold unchanged. It does not add or modify a Metal kernel.

The PR also adds a reproducible paired benchmark for the row-dispatch boundary.

Why

On an M1 Max with affine 4-bit/group-64 weights at K=5376, N=8192, the existing six-row matrix dispatch was consistently slower than a QMV control. Five independent 40-trial baseline runs measured chunked-QMV/direct ratios:

0.8717, 0.8930, 0.8947, 0.8767, 0.9152
median: 0.8930

The five-row same-route calibration median was 1.0052, which bounds local paired timing bias.

A final 100-trial run produced:

                 M=5 direct/chunk       M=6 direct/chunk
main             1.4492 / 1.4515 ms     1.9305 / 1.6750 ms
this patch       1.4501 / 1.4479 ms     1.6987 / 1.7290 ms

The direct six-row primitive decreased by about 12% in this isolated run. Across seven 200-trial runs, the baseline six-row chunk/direct median was 0.8807; after the threshold change it was 1.0118, showing the direct full-row QMV route aligns with the QMV control.

These are isolated primitive measurements, not end-to-end model or TTFT claims.

Reproduction

python benchmarks/python/qmm_dispatch_boundary.py \
  --rows 5 6 8 12 --warmups 10 --trials 100 --seed 61

Validation

• source build: pass
• python/tests/test_quantized.py: 51 passed, 1 skipped
• pre-commit on changed files: pass
• exact-shape direct/chunked numeric comparisons: pass (rtol=0.01, atol=0.06)
• independent review: pass

The full Python suite ran 943 tests; 4 unrelated test_fast_sdpa.py::test_sdpa_vector_head_dim_512 cases fail on this M1 Max because the kernels request 1024 threads while the device reports 640/832 maximum. The same failure reproduces on an unmodified build of the base revision.