feat(inference): comprehensive kernel micro-benchmarks + regression gate

[ohdearquant]

Motivation

The lm_head regression (#151) hid for months because we had no per-kernel timing — only E2E bench_decode_ab. Need a systematic micro-benchmark suite that catches per-kernel regressions before they compound.

Scope

Micro-benchmark binary (bench_kernel_isolate)

• Individual kernel timing: gemv_q8, gemv_decode_wide_f16, rms_norm, partial_rope, decode_attention (flash partial + reduce), silu_mul_fused, conv1d_silu, gdn_recurrence, fused_residual_add_norm, copy/add ops
• Each kernel: independent command buffer, 1000 iterations, median + p95 reported
• Full-layer composite (conv1d_silu + gdn_recurrence pipelined in single CB)
• Sum-of-parts vs E2E comparison to detect pipeline stalls

Regression gate

• CI job that runs micro-benchmarks on aarch64-macos (Apple Silicon)
• Per-kernel baseline stored on perf-baselines branch

• 10% regression on any kernel blocks merge (configurable threshold)

• Integrates with existing bench-regression.yml infrastructure from ADR-058

Tracking

• Baseline JSON committed per-platform
• Historical trend data for kernel-level performance over time

Acceptance

• bench_kernel_isolate binary compiles and runs on macOS
• Reports per-kernel timing with median/p95/stdev
• CI gate integrated with existing perf-baselines workflow
• Documents which kernels map to which model operations