ConcatDataset allows Packed sub datasets

[pbontrager]

Currently we don’t allow data packing with concat dataset. We can discuss enabling that, but in the meantime there’s a big in recipes wheee a user can set packed =True to each dataset in the list . This will individually pack each one but not use the right collate function, also, all of the datasets in the list should probably be packed together. To fix this for now, we can have ConcatDataset check that there are no packed datasets in the list.

[krammnic]

I see that this issue without "community welcome" label, but because it became little bit stale here are few thoughts. Probably something like:

for dataset in self._datasets:
    if isinstance(dataset, PackedDataset):
        raise ...

While initializing ConcatDataset will solve it, but knowing that we can import PackedDataset to ConcatDataset due to circular import, probably the better idea is to do something from this list:

1. Introduce some is_packed: bool property
2. Add utility to check this condition in utils(looks like bad idea)

So final solution will consist of:

1. Do some of the checks described above
2. Add unit test
3. Add some information in docs?

If you not mind, I might try to do a PR.

[RdoubleA]

There shouldn't be a circular import if ConcatDataset imports PackedDataset, as _packed.py does not currently depend on _concat.py. Would prefer doing the check within ConcatDataset, otherwise you would have to update all recipes.

[krammnic]

Ok, will recheck this, as I see torchtune does not have any custom exceptions, so I assume that raising some ValueError will be fine. Will create PR soon for this

[joecummings]

Closed by #1796