feat: add linfa-residual-sequence crate

[Feiyang472]

Hi! Wanted to float the idea of a small crate for chaining models together additively, fitting each model with the residuals of the previous one, like below

LinearRegression::default()
    .stack_with(some_svm)
    .stack_with(LinearRegression::default())
    .fit(&dataset)?

feels like a natural ergonomic addition which could sit as a crate in linfa.

Happy to hear if this belongs somewhere else (folded into an existing crate, or not at all).

[codecov]

Codecov Report

❌ Patch coverage is 88.46154% with 6 lines in your changes missing coverage. Please review.
✅ Project coverage is 76.76%. Comparing base (b1f9ddb) to head (4a712e6).

Files with missing lines	Patch %	Lines
src/composing/residual_chain.rs	88.46%	6 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #430   +/-   ##
=======================================
  Coverage   76.76%   76.76%           
=======================================
  Files         105      106    +1     
  Lines        7359     7411   +52     
=======================================
+ Hits         5649     5689   +40     
- Misses       1710     1722   +12     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[relf]

Hi! It looks like a nice addition! I think you should move it in the linfa crate under the existing composing module which gathers composite models.

[Feiyang472]

Hi @relf, absolutely composing is the better place for this to be! I have moved it to composing/residual-sequence.

[relf]

Ok, I think you should take a look at the contribution guidelines. Specially the sections about the implementation of PredictInplace trait and the optional serde support. As there is no hyperparameters (at the moment?), I guess you can skip the ParamGuard pattern.

[Feiyang472]

could also add another method which just shrinks by F::one() (no shrinking). not sure how to name it

[relf]

Well, the .shrink_by(1.0) is noisy... With a conversion we could have:

let model = MeanParams.stack_with(MeanParams.into()));

otherwise you can add another method:

let model = MeanParams.stack_with_full(MeanParams);

It is better if the most common use case has the shorter name.
We could also have just stack(MeanParams) and stack_shrunk(MeanParams, 0.5).

[Feiyang472]

Hi @relf thanks for reviewing. I propose the two methods:

chain(corrector: C)
chain_shrunk(corrector: Shrunk<C, F>)

I went with chain over stack since "stack" is so strongly associated with the fundamental data structure that it could be confusing in a library context. chain also echoes the ResidualChain struct name you suggested.

I kept chain_shrunk accepting a pre-wrapped Shrunk<C, F> rather than an inline shrinkage value (as in chain_shrunk(corrector, 0.5)) so that a Shrunk corrector can be constructed once and reused across multiple pipelines.