Align bench delta formatting with ember-eslint-parser

[NullVoxPopuli-ai-agent]

Ports ember-eslint-parser's benchmark delta calculation/formatting, so both repos' bench tooling reads the same way.

What changed

• Delta classification (the faster/slower logic):
  • ⚪ within ±2% — noise floor
  • 🟢 faster by more than 2% (previously a 2–5% win rendered as 🟡 "within 5%")
  • 🟠 slower by 2–5% — "slightly slower"
  • 🔴 slower by 5%+
• Shared scripts/bench-utils.mjs: formatTime, deltaEmoji, parsePairs, readBenchJSON, and the legend now live in one module used by both format-bench-cli.mjs and format-bench-comment.mjs (previously duplicated, and their thresholds had already drifted: the CLI/comment used a 1% white band while ember-eslint-parser uses 2%).

Example output

   Benchmark    Control (p50)   Experiment (p50)       Δ
   ──────────   ─────────────   ────────────────   ─────
⚪ gts small        100.00 µs           99.00 µs   -1.0%
🟢 gts large          2.00 ms            1.93 ms   -3.5%
🟠 gjs medium       500.00 µs          515.00 µs   +3.0%
🔴 hbs big          800.00 µs          860.00 µs   +7.5%

🟢 faster · 🔴 slower · 🟠 slightly slower · ⚪ within 2%

Verified with a synthetic results file covering all four bands (CLI table + PR comment markdown); pnpm lint clean.

🤖 Generated with Claude Code