Problem

When using ml_task="multiclass_classification" and setting eval_metric="f1",
the XGBoost algorithm was internally defaulting to accuracy as its evaluation
metric instead of respecting the user's choice.

This happened because the xgboost_eval_metric() function in
supervised/algorithms/xgboost.py only handled the logloss → mlogloss
mapping for multiclass, leaving f1 and accuracy without explicit routing
as custom metrics. As a result, XGBoost did not recognize f1 as a native
metric for multiclass tasks and silently fell back to accuracy.

Fix

Added an explicit elif branch in xgboost_eval_metric() to handle f1
and accuracy as custom metrics when ml_task is multiclass_classification,
consistent with how lightgbm_eval_metric() already handles this case in
supervised/algorithms/lightgbm.py.

Changes

• supervised/algorithms/xgboost.py: updated xgboost_eval_metric() to
  explicitly route f1 and accuracy as custom metrics for multiclass.
• tests/tests_fairness/test_multi_class_classification.py: added
  test_with_f1_metric() to cover the multiclass case with eval_metric="f1"
  and sensitive features.