[BUG] Performance Metric: multilevel="raw_values" with either multioutput="uniform_average" or custom weights for multioutput

[kdekker-private]

Describe the bug
Performance metric classes do not work when using the combination multilevel="raw_values" with either multioutput="uniform_average" or custom weights for multioutput.

To Reproduce

Using MeanAbsoluteError, but the error does not remain to this class.

import numpy as np
from sktime.performance_metrics.forecasting import MeanAbsoluteError
y_true = np.array([[0.5, 1], [-1, 1], [7, -6]])
y_pred = np.array([[0, 2], [-1, 2], [8, -5]])
mae = MeanAbsoluteError(multilevel="raw_values", multioutput="uniform_average")
mae(y_true, y_pred)
mae_custom_multitoutput_weights = MeanAbsoluteError(multilevel="raw_values", multioutput=[0.4,0.6])
mae_custom_multitoutput_weights(y_true, y_pred)

Output:

ValueError: DataFrame constructor not properly called!

Expected behavior
Both the output of mae and mae_custom_multitoutput_weights with custom multioutput weights should not throw an error.

Additional context
It seems like the wrapper class around the inner metric function fails to coerse the evaluations in the expected output format.

Versions
0.28.0

[fkiraly]

confirmed on current main, windows, python 3.11.

Note to devs: we need to check why this combination of inputs is not tested.

[fkiraly]

I am noting that the data in question has no hierarchy levels, as it is a "plain" series data container.

Still, the metric should not crash, so the coercion should be fixed.