Add streaming, dynamic, and smoothing regression (PassiveAggressive, LmDynamic, LOWESS)

[sipemu]

Three estimators that don't fit the classical batch-OLS pattern are available but unexposed.

Scope

Method	Rust path	Status in 0.5.6	Description
Passive-Aggressive	anofox_regression::solvers::PassiveAggressiveRegressor	new in 0.5.6	Online streaming regressor (PA-I / PA-II variants) with epsilon-insensitive loss. Same family as RLS but for non-quadratic loss; tunable update aggressiveness.
Dynamic LM	anofox_regression::solvers::lm_dynamic::*	pre-existing, unexposed	Time-varying-coefficient linear model with LOWESS weight smoothing. The natural "rolling regression" companion to OLS over time windows.
LOWESS	anofox_regression::solvers::lowess::*	pre-existing, unexposed	Locally weighted scatterplot smoothing — nonparametric smoother for visualization or feature engineering.

Suggested user-facing names

• passive_aggressive_fit_agg, passive_aggressive_fit_predict_by — primary surface is the streaming aggregate; emphasize compatibility with OVER (ORDER BY t).
• lm_dynamic_fit, lm_dynamic_fit_agg, lm_dynamic_fit_predict_by — likely takes a bandwidth / kernel option.
• lowess as a scalar/table macro (no fit/predict split — it returns a smoothed value per input row).

Prerequisite

Bump Cargo.toml: anofox-regression = "0.5.6" for PassiveAggressive. LmDynamic and LOWESS work against the current pin.

Tasks

• Bump dependency.
• Decide the right ergonomics for time-varying methods — window function vs aggregate vs table macro.
• FFI + C++ wrappers.
• SQL tests including a step-change time series for lm_dynamic, a streaming-update fixture for PassiveAggressive (verify behavior under repeated *_fit_agg over time-ordered partitions), and a noisy-scatter fixture for LOWESS.
• README updates — likely a new section since these don't fit the linear-models table cleanly.