benchmarks: add Apple I8MM affine-QMM evidence harness

[PhilipJohnBasile]

Summary

This draft adds an opt-in Apple arm64 evidence harness for comparing MLX affine 8-bit QMM against two dynamic-int8 activation controls:

• a U8×I8 NEON dot-product implementation
• a 2×2 U8×I8 I8MM implementation using vusmmlaq_s32

It follows the closed #4197 discussion and is intended only to inform a possible future ARM path alongside Highway work in #3019.

Scope

This draft:

• does not alter MLX production dispatch
• is excluded from normal builds unless explicitly enabled
• records activation preparation separately from same-preparation compute
• separates numerical-correctness runs from timing-only runs
• embeds the candidate source SHA256 and rejects a mismatched candidate
• records MLX binary/source identities as unbound provenance
• permanently reports throughput: null

It is not evidence of a public speedup, model-level quality, production readiness, Highway integration, or MLX source/binary provenance.

Why this shape

The earlier microbenchmark compared against a scalar stand-in and mixed activation-reuse gains with I8MM gains. This harness adds an actual MLX baseline, a same-preparation dot-product control, alternating measurement order, tail/edge coverage, and explicit fail-closed claim boundaries.

Validation

• fresh opt-in CMake configure and target build
• source-identity check and mismatch rejection
• normal self-test passed
• ASan and UBSan self-test passed
• diagnostic M=1/2/3 comparison passed
• CMake, C++, Python, and diff formatting checks passed

Possible extraction path

If the evidence eventually supports production work, activation reuse, the I8MM tile, and ARM Highway/CMake dispatch should be developed and reviewed as separate changes.