[BUG] transformers that are univariate-to-multivariate break when calling inverse_transform

[fkiraly]

Reported by @guimalo on discord, reproducing code:

from sktime.transformations.series.detrend import STLTransformer
from sktime.datasets import load_airline

X = load_airline()
t = STLTransformer(return_components=True)
Xt = t.fit_transform(X)
t.inverse_transform(Xt)

This should simply give back a series similar to X.

The breakage occurs as X has the univariate-only tag set, which will trigger inverse_transform to always vectorize over columns.
The solution should be that t vectorizes only over rows in this case.

That would solve the case above, but not a similar case where t had broadcast to a multivariate X in fit_transform.