Epic: Regularized GLMs and optimizers — SGDClassifier, L1 / ElasticNet logistic

[Mec-iS]

Tracking issue for linear-model solver work. Do not start code before an RFC is agreed here.

Member issues

• Implement SGDClassifier #62 — SGDClassifier
• Add L1 and Elastic Net regularization terms to the Logistic Regression #63 — L1 and Elastic Net regularization for LogisticRegression
• Add new solver that supports L1 penalty to Logistic Regression #70 — new solver with L1 penalty support for Logistic Regression

Why grouped

#63/#70 are the same capability seen from two angles (penalty term vs solver). #62 needs an optimizer abstraction that overlaps both. Decide the solver/penalty trait shape once, then implement members against it.

Current state (main @ a955334)

• LogisticRegression supports L2 only via LBFGS (src/linear/logistic_regression.rs).
• ElasticNet exists but is linear-regression only (src/linear/elastic_net.rs).

Constraints (repo conventions)

• Pure Rust, no macros; zero-copy iteration; numbers/ traits only.
• Keep the sklearn-inspired API surface consistent with existing linear models (getters: coefficients(), intercept()).
• Consider no_std impact (v0.7 roadmap) when choosing solver dependencies — none are wanted.

RFC requirements

Post a short RFC comment before any PR: penalty/solver trait sketch, convergence criteria, serde parameters, no_std impact.

Acceptance criteria

• RFC agreed by a maintainer
• Members implemented or explicitly descoped
• CHANGELOG updated