Add generic ResidualChain composing method (#430)

* feat: add linfa-residual-sequence crate

Implements ResidualSequence Struct and StackWith trait for composing regression
models in a boosting / residual-stacking pattern. The second (and any
further) model trains on the residuals left by the previous one;
predictions are summed.

Docs and tests were written with AI assistance.

* remove doc link

* move to composing/ module in linfa main crate

* update docs

* implement PredictInplace instead

* remove unused param error

* use one struct to implement stacking

* add deep chain test

* Rename to  ResidualChain
Implement Shrinkage
implement paramguard for shrinkage

* satisfy zola

* can only shrink by if target has the same float type

* work with predict inplace only

* zola fix

* add link in docs

* simplify comparison

* rename to residual_chain as consistent with struct

* implement copy trait

* add method `chain` which just chains self with an unshrunk corrector.
rename stack_with -> chain_shrunk

* add doc

Author: Feiyang472

Cargo.toml

@@ -63,6 +63,8 @@ linfa-datasets = { path = "datasets", features = [
     "diabetes",
     "generate",
 ] }
+linfa-linear = { path = "algorithms/linfa-linear" }
+linfa-svm = { path = "algorithms/linfa-svm" }
 statrs = "0.18"
 
 [target.'cfg(not(windows))'.dependencies]

src/composing/mod.rs

@@ -1,12 +1,15 @@
 //! Composition models
 //!
-//! This module contains three composition models:
+//! This module contains four composition models:
 //!  * `MultiClassModel`: combine multiple binary decision models to a single multi-class model
 //!  * `MultiTargetModel`: combine multiple univariate models to a single multi-target model
 //!  * `Platt`: calibrate a classifier (i.e. SVC) to predicted posterior probabilities
+//!  * `ResidualChain`: fit models sequentially on the residuals of the previous one
+//!    (forward stagewise additive modeling / L2Boosting); see [`residual_chain::Stagewise`]
 mod multi_class_model;
 mod multi_target_model;
 pub mod platt_scaling;
+pub mod residual_chain;
 
 pub use multi_class_model::MultiClassModel;
 pub use multi_target_model::MultiTargetModel;