Improve is_probabilistic_classifier (fix lighting tests)

Try to look inside estimators that wrap other estimators:
predict_proba can be a property in wrapped estimator but a method
in wrapper estimator, so the AttributeError will be raised in
runtime. But we don't want to catch it in runtime because it might
be unrelated. This fixes a failure that appeared after
scikit-learn-contrib/lightning#98 merge
(see https://travis-ci.org/TeamHG-Memex/eli5/jobs/172586529)

eli5/sklearn/utils.py

@@ -1,5 +1,6 @@
 # -*- coding: utf-8 -*-
 import numpy as np
+from sklearn.base import MetaEstimatorMixin
 
 
 def is_multiclass_classifier(clf):
@@ -19,7 +20,11 @@ def is_multitarget_regressor(clf):
 
 def is_probabilistic_classifier(clf):
     """ Return True if a classifier can return probabilities """
-    return hasattr(clf, 'predict_proba')
+    if not hasattr(clf, 'predict_proba'):
+        return False
+    if isinstance(clf, MetaEstimatorMixin) and hasattr(clf, 'estimator'):
+        return hasattr(clf.estimator, 'predict_proba')
+    return True
 
 
 def has_intercept(estimator):