Add CodSpeed continuous benchmarking

[codspeed-hq]

This sets up continuous performance measurement for pymath with CodSpeed.

What was added

Benchmark harness (divan via codspeed-divan-compat)

codspeed-divan-compat is added as a dev-dependency renamed to divan, so the benchmarks are plain divan benchmarks: they run standalone locally and are instrumented automatically when executed through CodSpeed.

Three benchmark suites in benches/:

• math.rs — real-valued functions, grouped by source module:
  • trigonometric: sin, cos, tan, asin, acos, atan, atan2, sinh, cosh, tanh, asinh, acosh, atanh
  • exponential: exp, exp2, expm1, log (natural and with an explicit base), log2, log10, log1p, sqrt, cbrt, pow
  • gamma: erf, erfc (including the |x| > 1.5 continued-fraction branch), gamma for positive and negative arguments, lgamma
  • misc: fabs, floor/ceil/trunc, frexp/modf, fmod, remainder, fma, nextafter with steps, ldexp/ulp, isclose
  • aggregate: fsum, dist, sumprod, sumprod_int, prod at 8 / 128 / 4096 elements to cover both the short and long reduction paths
• cmath.rs — complex functions (exp, log, log10, sqrt, the six trigonometric and six hyperbolic functions plus their inverses, abs/phase, polar/rect) over a spiral of inputs spanning all four quadrants
• integer.rs — big-integer functions (factorial, comb, perm, isqrt, gcd, lcm, log_bigint) at several magnitudes, gated behind required-features = ["num-bigint"]

Each scalar benchmark sweeps a fixed set of representative finite inputs inside the measured closure, so results reflect the function under test rather than harness overhead, and the input domains are chosen to stay in the valid range of each function.

CI workflow — .github/workflows/codspeed.yml runs the benchmarks in CPU simulation mode on ubuntu-latest (same runner as the existing Rust workflow), on pushes to main, on pull requests, and on workflow_dispatch so CodSpeed can backfill baseline data. It authenticates with OIDC (id-token: write), builds with cargo codspeed build --features num-bigint so all three suites are included, and Python is set up because the crate's dev-dependencies include pyo3 with auto-initialize.

README — added the CodSpeed badge.

Validation

All three suites were built and executed locally with codspeed run --mode simulation -- cargo codspeed run; the full set of benchmarks completed and reported results. cargo clippy --benches --features num-bigint and cargo fmt are clean.

Next steps

• Merge this PR so the first main run establishes the performance baseline.
• Subsequent pull requests will get a CodSpeed report comparing against that baseline.
• New functions ported from CPython can be covered by adding them to the corresponding module in benches/.

[codspeed-hq]

Congrats! CodSpeed is installed 🎉

🆕 95 new benchmarks were detected.

You will start to see performance impacts in the reports once the benchmarks are run from your default branch.

Detected benchmarks

• abs_phase: 32 µs
• acos: 70 µs
• acosh: 71.2 µs
• asin: 74 µs
• asinh: 75.6 µs
• atan: 45.3 µs
• atanh: 44.8 µs
• cos: 56.8 µs
• cosh: 56.2 µs
• exp: 38.6 µs
• log_base: 71 µs
• log_natural: 45.5 µs
• log10: 41.9 µs
• polar_rect: 55 µs
• sin: 56.8 µs
• sinh: 56.4 µs
• sqrt: 18 µs
• tan: 51.2 µs
• tanh: 51 µs
• comb[(30, 15)]: 1.3 µs
• ...

ℹ️ Only the first 20 benchmarks are displayed. Go to the app to view all benchmarks.

---