FIX: resolve sklearn 1.9 compatibility for WeightedLassoCV n_alphas deprecation (#1032 )

[genrichez]

Summary

Fixes InvalidParameterError and FutureWarning when using EconML with scikit-learn >= 1.7 (where n_alphas is deprecated) and >= 1.9 (where it will be removed).

Closes #1032

Problem

In sklearn 1.7+, LassoCV and MultiTaskLassoCV deprecated the n_alphas parameter in favor of passing an int directly to alphas. The existing version check in WeightedLassoCV.__init__ correctly passes alphas=n_alphas to super().__init__(), but two issues remained:

1. self.alphas was set to None (the default alphas kwarg value) instead of the effective int value, causing InvalidParameterError: The 'alphas' parameter must be an int or array-like. Got None instead at fit time.

2. n_alphas leaked into get_params() output. When sklearn's LinearModelCV.fit() calls self.get_params() to build path_params, n_alphas was passed to lasso_path(), triggering FutureWarning: 'n_alphas' was deprecated in 1.9 and will be removed in 1.11.

Fix

• Set self.alphas to the effective value (alphas if provided, else n_alphas as int) in the >= 1.7 branch
• Override get_params() to exclude n_alphas from the returned dict on sklearn >= 1.7, preventing it from leaking into lasso_path(**path_params)
• Apply same fix to WeightedMultiTaskLassoCV

Testing

Verified with sklearn 1.9.0:

Test	Result
Import econml.orf (triggers WeightedLassoCV init)	PASS
WeightedLassoCV.fit() with default n_alphas=100	PASS
WeightedLassoCV.fit() with n_alphas=1 (edge)	PASS
WeightedLassoCV.fit() with n_alphas=1000 (large)	PASS
WeightedLassoCV.fit() with explicit alphas array	PASS
Both n_alphas and alphas provided (alphas takes precedence)	PASS
WeightedMultiTaskLassoCV.fit()	PASS
WeightedLassoCVWrapper single/multi-target	PASS
WeightedLassoCVWrapper with sample_weight	PASS
DebiasedLasso (uses WeightedLassoCV internally)	PASS
LinearDML end-to-end pipeline	PASS
sklearn.clone() compatibility	PASS
set_params() round-trip	PASS
Pickle serialization round-trip	PASS
Zero FutureWarning about n_alphas	PASS
Alpha grid size matches n_alphas (10, 50, 100, 200)	PASS
WeightedLassoCV == sklearn LassoCV when no weights (coef diff=0)	PASS
Sample weights correctly affect coefficients	PASS
DebiasedLasso CI coverage ~95%	PASS

Backward Compatibility

• WeightedLassoCV(n_alphas=50) still works on all sklearn versions
• WeightedLassoCV(alphas=[0.1, 0.01]) still works
• get_params() / set_params() / clone() all work correctly
• No behavior change on sklearn < 1.7