Fix classification predictions with non-numeric class labels - #830

Closes #434.

predict() always casts classification results to np.int_, which blows up for any classifier trained on non-numeric class labels. The iris dataset's species names are the obvious example:

ValueError: invalid literal for int() with base 10: 'setosa'

Fix tries the np.int_ cast first, then falls back to letting numpy infer the dtype if that cast raises. Kept the try-first order on purpose: turns out Elasticsearch can return numeric class labels as numeric-looking strings (depending on how the model got serialized), so a plain "is this a string" check would have broken the existing int-labeled classifier tests. Found that out the hard way while testing this against a real cluster.

Verified against a real Elasticsearch 9.2.0 instance, not mocked:

• Trained a RandomForestClassifier on iris with the actual string species names, imported it, watched predict() throw the ValueError above pre-fix, then confirmed it returns string predictions matching sklearn's own output post-fix.
• Ran the existing test_decision_tree_classifier cases (all 4 parametrizations) to make sure numeric-labeled classifiers still come back as np.int_. No regression there.
• Added test_decision_tree_classifier_string_class_labels next to them, same pattern, covering the case from the issue.

Also fixed the predict() docstring, it still claimed results are always float or int.