Remove redundant learning curve implementation paths

[bruAristimunha]

Summary

This PR unifies the learning curve implementation by removing redundant code paths and consolidating all learning curve functionality into the LearningCurveSplitter class. This makes learning curves available across all evaluation types (WithinSession, CrossSession, CrossSubject) through a consistent API.

Changes

New Features

• Add cv_class and cv_kwargs parameters to all evaluation classes for custom cross-validation strategies
• Implement LearningCurveSplitter as a dedicated sklearn-compatible cross-validator for learning curves
• Learning curve results now automatically include "data_size" and "permutation" metadata columns

API Changes

• Breaking: Remove data_size and n_perms parameters from WithinSessionEvaluation
• Use cv_class=LearningCurveSplitter with cv_kwargs=dict(data_size=..., n_perms=...) instead

Code Health

• Centralize CV resolution in BaseEvaluation with new _resolve_cv() method
• Remove redundant methods from WithinSessionEvaluation:
  • get_data_size_subsets()
  • score_explicit()
  • _evaluate_learning_curve()
• Simplify evaluate() to always use unified _evaluate() path

Other Improvements

• Add single-class safeguard for LearningCurveSplitter (skip splits where training set collapses to single class)
• Fix ArrowStringArray shuffle warnings by converting to numpy arrays
• Update all examples and tests to use new API

Migration Guide

Before:

evaluation = WithinSessionEvaluation(
    paradigm=paradigm,
    datasets=datasets,
    data_size={"policy": "ratio", "value": [0.2, 0.4, 0.6, 0.8, 1.0]},
    n_perms=[5, 5, 5, 5, 1],
)

After:

from moabb.evaluations.splitters import LearningCurveSplitter

evaluation = WithinSessionEvaluation(
    paradigm=paradigm,
    datasets=datasets,
    cv_class=LearningCurveSplitter,
    cv_kwargs=dict(
        data_size={"policy": "ratio", "value": [0.2, 0.4, 0.6, 0.8, 1.0]},
        n_perms=[5, 5, 5, 5, 1],
    ),
)

Test Plan

• All existing tests pass
• Learning curve examples work with new API
• Documentation updated in whats_new.rst

Closes #962