Fix regressor fit crash when sklearn transform_output is "pandas"

[fus3r]

Why

TabFMRegressor.fit crashes for any user whose sklearn output config is set to
pandas globally:

import sklearn
sklearn.set_config(transform_output="pandas")
reg.fit(X_train, y_train)
# AttributeError: 'DataFrame' object has no attribute 'flatten'

The target StandardScaler is an internal transformer. Under
transform_output="pandas" its fit_transform returns a DataFrame, and the
next line calls .flatten() on it, which only exists on ndarrays. That is
issue #58 (the reporter's traceback shows the matching FunctionTransformer
"pandas" warning, so the config was active in their environment).

The classifier is not affected (it does not scale the target). predict is not
affected either: inverse_transform output is not controlled by that config.

What

Pin the internal target scaler to ndarray output with
set_output(transform="default"), so the regressor ignores the user's global
config for its own internal transformer. This is the standard way for a library
to keep ndarray output when it consumes transformer results directly.

Added a regression test that fits the regressor inside a
transform_output="pandas" context and checks it does not crash.