Optimize Async poly interpolation hot path via Horner-form polynomial evaluation

[Copilot]

Issue #135 identified many optimization ideas in Async::new_poly; the high-value, low-risk part was reducing arithmetic cost in polynomial evaluation. This PR implements only that part by replacing explicit power expansion with Horner-form evaluation in the poly interpolation kernels.

• What changed

  • src/asynchro_fast.rs: polynomial kernel rewrite
    • Updated interp_septic, interp_quintic, and interp_cubic to evaluate the same polynomials using Horner form.
    • Removed intermediate power terms (x2, x3, ...), reducing multiply count and tightening the inner interpolation path.

• Scope intentionally kept narrow

  • No structural refactors
    • Did not introduce scratch-buffer reuse, loop reordering, or API/behavior changes from the broader issue discussion.
    • Kept this PR focused on arithmetic optimization in the existing implementation shape.

• Example (pattern used)

// before
let x2 = x * x;
let x3 = x2 * x;
a0 + a1 * x + a2 * x2 + a3 * x3

// after
a0 + x * (a1 + x * (a2 + x * a3))