Add a dev-only nanobind Python layer and the beta distribution's entropy

[NAThompson]

nanobind/ bindings (not built by CI); entropy(beta_distribution) via Stirling and zeta series; libeta, ligamma and lbeta return an exact -inf instead of raising; x87 tolerance in test_lanczos_smoothing.

Closes #1508

entropy(beta_distribution)

The textbook form log B(α, β) − (α−1)ψ(α) − (β−1)ψ(β) + (α+β−2)ψ(α+β) cancels catastrophically. For large α or β the terms grow like α log α while H only grows like log α, and near α = β = 1 the terms are first order while H is second order.

• α or β ≥ 10 (scaled with precision): Stirling remainder series cancel the large parts exactly, giving H = ½ log(2πe αβ/s³) − 1/(2α) − 1/(2β) + 1/s + r(α) + r(β) − r(s) + ν(s), with s = α + β.
• Both small: H = g(α) + g(β) + k(s), where g(x) = lgamma(x) − (x−1)ψ(x) and k(s) = (s−2)ψ(s) − lgamma(s) vanish to second order at 1 and 2. Near those points they come from zeta series.

Against mpmath at 40 digits on a 500 × 500 log grid, double precision, the median error is 0.63 ulps and the maximum 110. The maximum is at 1 ≲ α, β ≲ 10, where neither series applies. The textbook form reaches 5.3·10⁶ ulps.

cpp_bin_float_50 gives 0.03 to 21 eps at the test points.

🤖 Generated with Claude Code