perf: use median aggregation for reliable benchmark comparisons

[vm75]

Summary

Improve performance benchmark reliability by replacing arithmetic-mean aggregation with the median across repeated runs, reducing sensitivity to occasional CPU contention and background activity.

Changes

• Use median aggregation for throughput, p50, and p95 timing metrics.
• Increase local run-perf.py default from 5 to 9 runs while retaining 5 runs for CI and Makefile defaults.
• Normalize historical benchmark snapshots from raw measurements to ensure consistent mean-versus-median comparisons.
• Preserve existing CLI parameters, environment variables, CSV formats, raw measurements, and regression thresholds.
• Add regression tests for aggregation, CSV compatibility, and historical snapshot normalization.
• Update performance documentation.

Compatibility

No changes to the transliteration engine, public APIs, or native benchmark workloads. Existing CI and performance tooling remain compatible.

Validation

Added focused regression tests. Test execution and CI validation should be confirmed before merging.