Optimize univariate evaluation and power backends

[eliasjudin]

Summary

Univariate evaluation and power backends: replace current naive evaluation with a Horner-style implementation and replace the public raw power backend with exponentiation by squaring.

Keep the public Raw.eval₂, Raw.eval, CPolynomial.eval₂, and CPolynomial.eval entrypoints, and keep the public Raw.pow / Pow surface update the proof layer to the correct lawful theorem surface.

Correctness target

Prove that the Horner evaluator agrees with the existing univariate evaluation semantics and that the fast power backend agrees with canonical repeated multiplication at the lawful layer and on the canonical CPolynomial / QuotientCPolynomial surfaces.

In particular, re-establish the toPoly bridge theorems for evaluation and the lawful power theorems needed to preserve the existing canonical semiring behavior.

Motivation

Advances the phase 2 univariate performance track: improves the asymptotic cost of evaluation and exponentiation without changing the public behavior of the univariate API.