Binary risk control - multi risk

[allglc]

Context

• Add the possibility of multi risk control
• Keep unidimensional lambda (using thresholding on predict_proba)

Content

This PR includes (lists incrementally updated):

• Update typing of BinaryClassificationController

[Valentin-Laurent]

Good start 💪🏻

[Copilot]

Pull Request Overview

This PR adds support for multi-risk control to the BinaryClassificationController by updating type annotations and adding validation logic. It allows users to specify multiple risks and corresponding target levels while maintaining backward compatibility with single risk control.

• Updated type hints to accept lists of risks and target levels alongside single values
• Added validation to ensure risks and target levels lists have matching lengths
• Enhanced class documentation to describe multi-risk control functionality

---

_{Tip: Customize your code reviews with copilot-instructions.md. Create the file or learn how to get started.}

[Valentin-Laurent]

Reviewing the __init__ method and submethods.

Nice job!

Regarding typing, we could use a slightly stricter MyPy flag to forbid untyped functions (right now, only partially untyped function are forbidden). But if I remember, using that flag would cause a lot of typing to do in the existing codebase (not 100% sure though). So right now, let's try to not forget typing new code.

cc @GBrelurut

[Valentin-Laurent]

Reviewing method calibrate (except ltt_procedure).

[Valentin-Laurent]

This method is now more generic.

It is not directly unit tested (it is tested undirectly in TestBinaryClassificationControllerSetBestPredictParam).

We should probably add a multi-risk test case in that class (either in existing test or as a new one).

[allglc]

TestBinaryClassificationControllerSetBestPredictParam tests _set_best_predict_param which only calls this method with one risk (_best_predict_param_choice). The call for multi risk happens in calibrate. So I need to write a test in another class or a new unit test. On it!

[allglc]

done

[allglc]

My latest commit updates ltt_procedure for multi risk. Calls to the function and tests have been adapted. Input types and/or shapes have been modified.

I also replaced np.where by np.nonzero, which is recommended by numpy doc in the case when only the condition is provided.