[jolars]

This repo has no Lasso or Elastic Net benchmarks to compare the fit changes in smartcorelib/smartcore#480. This PR adds benchmarks for both estimators and instructions to compare eager and lazy normalization with a local smartcore checkout.

The Criterion benchmarks use fixed f64 data at four sizes: 1,024 × 32, 8,192 × 32, 65,536 × 32, and 1,024 × 128. Each size runs with row-major and column-major DenseMatrix inputs and a row-major ndarray input. Data creation occurs outside the timed loop. Both estimators use normalization and the same solver settings across runs.

The iai-callgrind benchmarks add three cases per estimator: normalized fits in both dense layouts and a row-major fit without normalization. Setup occurs outside the instruction count. Both CI jobs include the new benchmarks and continue to use the published smartcore release.

The regularized_fit example fits one model per process for peak-memory measurements with GNU time. The README explains how to run these measurements and save an eager Criterion baseline for comparison with a lazy build.

Validation:

• cargo bench --no-run
• cargo fmt --all -- --check
• cargo clippy --all-targets -- -D warnings
• All 24 Criterion cases passed in --test mode with smartcore 0.6.15 and with the local PR checkout at aaa79c9, both with and without smartcore/lazy-normalization.
• All six iai-callgrind cases ran under Valgrind. The history converter produced six entries with unique names.
• The memory example ran for both estimators with the published release and the local lazy build.