Storage policy's buffer doesn't keep the targets attribute

[HamedHemati]

All storage policies in Avalanche initialize their internal buffer with an empty dataset, i.e.:

self._buffer: AvalancheDataset = concat_datasets([])

When a new dataset is used to update the current buffer, self._buffer is concatenated with the new dataset, and since it's a concatenation of an empty dataset with another dataset , the targets are not transferred in the case of classification datasets.

Is this a bug or is it the expected behavior?

[AntonioCarta]

Definitely a bug. AvalancheDataset methods should ignore empty datasets when concatenating targets. This is probably a regression because I remember it used to work.