[python-package] fix misleading feature name warning on sklearn 1.6+ predict()

[FranciscoRMendes]

Summary

Fixes #6798.

Starting with scikit-learn 1.6, LGBMClassifier/LGBMRegressor/etc. emit a spurious warning during predict() when fitted on data without feature names (e.g. a numpy array):

UserWarning: X does not have valid feature names, but LGBMClassifier was fitted with feature names

This happens because LightGBM auto-generates feature names (Column_0, Column_1, ...) for such inputs and previously exposed them via feature_names_in_, causing sklearn to believe the model was fitted with named features.

Fix: feature_names_in_ now raises AttributeError when training data had no feature names, matching sklearn's own convention. Auto-generated names remain accessible via the LightGBM-specific feature_name_ property.

Changes

• python-package/lightgbm/sklearn.py: track _fitted_with_feature_names in fit(); gate feature_names_in_ on that flag
• tests/python_package_test/test_sklearn.py: update test_getting_feature_names_in_np_input to assert feature_names_in_ is absent for numpy input; add regression test test_no_spurious_feature_name_warning_on_np_predict

Test plan

• test_getting_feature_names_in_np_input — asserts feature_names_in_ is not set after numpy fit, feature_name_ still works
• test_no_spurious_feature_name_warning_on_np_predict — asserts no warnings raised during predict() on numpy input
• test_getting_feature_names_in_pd_input — unchanged; DataFrame input still exposes feature_names_in_
• Full test_sklearn.py suite: 483 passed, 0 failures