Fix regressor fit crash when sklearn output is globally set to pandas

[devYRPauli]

Problem

TabFMRegressor.fit scales the regression target with a StandardScaler and then calls .flatten() on the transformed result. When a user has sklearn.set_config(transform_output="pandas") set globally (common in analysis notebooks), fit_transform returns a pandas DataFrame, which has no .flatten() method, so fitting fails with:

AttributeError: 'DataFrame' object has no attribute 'flatten'

This is issue #58 (regression example not working).

Fix

Pin the target scaler to numpy output via set_output(transform="default"), so target scaling is unaffected by the user's global sklearn output configuration.

Note: only this site needs the change. The classifier's CategoricalOrdinalEncoder path is not affected (that encoder is never wrapped to pandas output), and inverse_transform calls are never wrapped either, so no other sites required a fix.

Testing

• Reproduced the AttributeError on the regressor fit path with transform_output="pandas" before the change, and confirmed fit completes cleanly after it.
• The non-JAX unit tests in classifier_and_regressor_test pass locally.

[fus3r]

Hi, I think this may overlap with #59. That PR also includes a regression test for the global pandas output configuration, which may be useful here.

[devYRPauli]

Thanks for pointing that out, and apologies for the duplicate - I should have checked the open PR list against #58 before sending this.

You are right that #59 covers this: same fix in the same place (pinning the y scaler with set_output(transform="default") before fit_transform().flatten()), and it also carries the regression test under sklearn.config_context(transform_output="pandas"), which mine does not. Yours is both earlier and more complete, so there is nothing here worth a second review.

Closing this in favor of #59.