refactor: extract DomainConstants struct

[MauroToscano]

Summary

• Introduces DomainConstants<F> in domain.rs — groups the four precomputed barycentric interpolation values (points, offset_pow_n, size_inv, offset_pow_n_inv) with a from_domain constructor
• Reduces get_trace_evaluations_from_lde from 10 parameters to 6, removing the #[allow(clippy::too_many_arguments)] suppression
• Replaces the 6-line inline computation block in round_3 (prover.rs) with DomainConstants::from_domain(domain)
• Simplifies the test from 8 lines of manual precomputation to one constructor call

Test plan

• All 124 stark lib tests pass (cargo test -p stark --lib)
• barycentric_trace_eval_matches_horner_trace_eval — verifies the refactored path matches Horner evaluation
• test_deep_poly_direct_2n_matches_interpolate_fft_extend — verifies DEEP poly direct 2N path

Pure refactor, no behavioral changes.

[claude]

Clean refactor. DomainConstants is a natural home for these four values, the removal of #[allow(clippy::too_many_arguments)] is a good signal it was the right call, and the test simplification validates the design.

Two minor issues flagged inline:

1. trace.rs:397 — local variable n_inv_g_n_inv is a stale name leftover from before the rename; the surrounding comment already uses the new size_inv/offset_pow_n_inv names.
2. domain.rs:34 — the expect message says "hence non-zero in the field" which is only true when char(F) does not divide n; the message should reflect the actual assumption rather than assert a consequence.

No security or correctness issues. The debug_assert_eq! on dc.points.len() is still present and sufficient for catching domain/constants mismatches in debug builds.

[github-actions]

Codex Code Review

No issues found in the PR diff.

I also attempted a targeted cargo test -p stark ..., but the environment blocked rustup from creating a temp file under /home/runner/.rustup due a read-only filesystem, so I could not verify tests locally. git diff --check passed.