feat: adaptive Simpson's quadrature (integrate_adaptive)

[bwyard]

Context

integrate_simpson uses fixed subdivisions. For functions with localized variation (spikes, discontinuities), adaptive refinement is needed to match SciPy/GSL accuracy.

Proposed API

pub fn integrate_adaptive(f: fn(f32) -> f32, a: f32, b: f32, tol: f32, max_depth: u32) -> f32

Recursive adaptive Simpson's with Richardson extrapolation. Subdivides panels where |S_fine - S_coarse| > 15 * tol. O(h^4) per panel.

Thesis alignment

• Pure: (f, a, b, tol, max_depth) -> f32
• ADVANCE-EXCEPTION: recursion bounded by max_depth
• No &mut in public API

Tests needed

• sin(x) over [0, π] = 2.0 within 1e-6
• More accurate than fixed Simpson at same evaluation count
• Handles discontinuities (step function)
• Deterministic