kimchi: merge witness commitment and interpolation into single pass

[dannywillems]

Summary

• Merge the witness column commitment and polynomial interpolation into a single parallel loop
• Eliminates the full witness array clone (~30 MB for domain size 2^16)
• Removes 15 redundant Evaluations constructions and 15 redundant iFFTs

Before

Two separate parallel loops:

1. witness.clone().into_par_iter() — clones all 15 columns, commits each
2. (0..COLUMNS).into_par_iter() — clones each column again, interpolates each

Total: 30 column copies, 30 iFFTs (15 for Lagrange-basis commit + 15 for interpolation)

After

One parallel loop using witness.par_iter() (borrows):

1. Clone column once for from_vec_and_domain
2. Commit via commit_evaluations_non_hiding (borrows &Evaluations)
3. Interpolate in-place via interpolate(self) (arkworks ifft_in_place)

Total: 15 column copies, 15 iFFTs

Benchmark results

cargo bench --bench proof_criterion -p kimchi (Apple M-series, sample size 10):

Benchmark	master	optimized	Change
1014 gates (SRS 2^16)	671.10 ms	674.64 ms	+0.53% (p=0.50, noise)
32758 gates (SRS 2^16)	1.3233 s	1.2669 s	-4.26% (p=0.00)

Improvement scales with circuit size — larger witness columns benefit more from eliminating redundant clones and iFFTs.

Test plan

• cargo clippy -p kimchi --all-targets --all-features passes clean
• Benchmark shows statistically significant improvement at scale
• CI passes

[dannywillems]

This was vibe-coded, PR opened by Claude automatically.