bench: add benchstat-based comparison workflow for branch vs baseline

[autholykos]

Summary

Add a lightweight benchmark comparison workflow using benchstat so contributors can compare benchmark output against a baseline instead of reading raw numbers manually.

Why this matters

• raw go test -bench output is hard to review consistently
• we have benchmarks, but not a reliable comparison story
• before and after perf claims should be measurable, especially for indexing and targeted analysis work

Proposed shape

• add a script or make target that:
  • captures a baseline benchmark output
  • captures a candidate benchmark output
  • runs benchstat on the two files
• document a simple workflow for comparing:
  • main vs current branch
  • before vs after a local optimization
• keep this opt-in; no CI gate required initially

Nice to have

• support corpus or profile labels in output
• support both bootstrap and larger benchmark corpora
• optionally save outputs under a temp or ignored directory

Acceptance criteria

• contributors can run one documented command to compare two benchmark runs
• results include clear per-benchmark deltas
• docs explain how to use the comparison workflow for perf-sensitive PRs
• the workflow works with the existing benchmark suite without changing default CI