ci: add performance regression detection for exact-arithmetic benchmarks

[acgetchell]

Summary

Add a GitHub Actions workflow that detects performance regressions in the exact-arithmetic benchmark suite on PRs and pushes to main.

Current State

• Exact-arithmetic benchmarks exist in benches/exact.rs (D=2–5, near-singular 3x3)
• scripts/bench_compare.py can compare Criterion baselines and generate markdown tables
• Baselines are saved locally via just bench-save-baseline <TAG> but there is no CI integration
• The delaunay project has a working implementation (.github/workflows/benchmarks.yml + scripts/benchmark_utils.py) that can serve as a reference

Proposed Changes

GitHub Actions workflow (.github/workflows/benchmarks.yml)

1. Trigger: on push to main and on PRs
2. Baseline strategy: save a Criterion baseline on main merges as a GitHub Actions artifact; download it on PR runs for comparison
3. Regression detection: run cargo bench --features bench,exact --bench exact, compare against the main baseline, flag regressions above a configurable threshold (e.g., 7.5%)
4. Reporting: post a summary comment on the PR with the comparison table (reuse bench_compare.py output or Criterion's built-in comparison)

Baseline artifact management

• Save baseline artifacts on main branch pushes (retain for ~30 days)
• PRs download the latest main baseline for comparison
• Optionally save tagged-release baselines with longer retention (90 days)

Considerations

• Hardware consistency: GitHub Actions runners have variable performance; may need a noise threshold higher than local benchmarks
• Runtime: exact benchmarks currently take ~30s with --quick; full run ~2–3 min — acceptable for CI
• False positives: the threshold should be tuned after a few weeks of data; start conservative (e.g., 10%) and tighten later

Benefits

• Catch performance regressions before merge
• Automated tracking — no manual benchmark runs needed for regression detection
• Complements the existing local just bench-compare workflow

Implementation Notes

• Reference: ../delaunay/.github/workflows/benchmarks.yml and ../delaunay/scripts/benchmark_utils.py
• Consider using github-action-benchmark or Criterion's built-in --baseline comparison
• The workflow should not block PRs on regression (warning only) until thresholds are tuned
• Related: perf: custom f64 → BigRational conversion via mantissa/exponent decomposition #63 (f64→BigRational optimisation), perf: stack-backed storage for exact arithmetic kernels #62 (stack-backed storage)

[acgetchell]

Closed by #74