[Perf] Expand apples-to-apples kernel suite and scorecard

[Simonsbs]

Goal

Improve runtime competitiveness beyond single-kernel parity by expanding apples-to-apples coverage.

Scope

Add at least 8 additional kernel shapes, each with L0+C equivalents and a shared harness:

1. integer arithmetic chain
2. bitwise-heavy kernel
3. branch-heavy kernel
4. memory load/store roundtrip
5. pointer arithmetic loop
6. function call chain
7. mixed arithmetic+branch kernel
8. small struct/aggregate pass

Acceptance

• All kernels benchmarked in CI.
• Median-of-N reporting per kernel.
• Results page includes per-kernel winner and geometric mean.

[Simonsbs]

Implemented first major slice for #2:

• Reworked tests/benchmark_apples_to_apples.sh from single-kernel to multi-kernel matrix (9 kernels):
  add, sub, mul, and, xor, cbr-select, memory-roundtrip, call-add, sum6-sysv.
• Added per-kernel build/runtime ratios (L0/GCC).
• Added aggregate geometric-mean summary for build and runtime ratios.
• Regenerated docs/PERFORMANCE_COMPARISON_APPLES_TO_APPLES.md with expanded scorecard.

Current snapshot (local run):

• Geometric mean build ratio (L0/GCC): 17.5596
• Geometric mean runtime ratio (L0/GCC): 0.9988 (near parity)

Next:

• Add additional kernel diversity for branch-heavy and mixed arithmetic+branch stress paths.
• Add confidence interval metadata and repeatability controls as a follow-up under reproducibility backlog.