Fix flaky test_logistic_regression_weighting test

[csadorf]

Summary

Refactors test_logistic_regression_weighting to use Hypothesis for property-based testing, fixing flakiness in multiclass logistic regression with class weighting.

Closes #7298

Problem

The test_logistic_regression_weighting[multiclass-3-class_weight] test was failing intermittently. The root causes:

1. Insufficient iterations: Some weight combinations needed more iterations to converge
2. Fixed tolerances: Didn't account for varying problem difficulty (weights, number of classes)
3. Normalization bug: Missing keepdims=True in multiclass coefficient normalization

Changes

Converted to Hypothesis-based testing:

• Uses standard_classification_datasets() with use_sample_weight and class_weight_option parameters
• Adds 4 explicit @example decorators for critical scenarios (including the original failing case)

Improved robustness:

• Dynamic max_iter: 500 → 1000 for problems combining sample weights and class weights
• Adaptive tolerances:
  • Easy problems (no weights, ≤3 classes): unit_tol=0.04, total_tol=0.08
  • Hard problems (with weights or >3 classes): unit_tol=0.25, total_tol=0.40

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[csadorf]

/ok to test 557fee1

[csadorf]

/ok to test d849ced

[csadorf]

/ok to test e86e662

[csadorf]

/merge