Fix wrapped packing duplicating data when the input table is sliced - #6670

What does this PR do?

Fixes #6669

_pack_wrapped sizes the output from the array's sliced view but reads from the raw child buffer, so when the input table is a slice it re-reads from the start of the buffer instead of from the slice. datasets.map(batched=True) passes each
batch as a slice, so every batch after the first can come back holding data from the beginning of the dataset. Row and token counts are unchanged, so nothing errors.

ds = Dataset.from_dict({"input_ids": [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12]]})
pack_dataset(ds, seq_length=4, strategy="wrapped", map_kwargs={"batch_size": 3})["input_ids"]
# before: [[1, 2, 3, 4], [5, 6], [1, 2, 3, 4], [5, 6]]
# after:  [[1, 2, 3, 4], [5, 6], [7, 8, 9, 10], [11, 12]]

The correction was already computed into a local, but #5189 rewrote the loop as a list comprehension over all columns and passed column.values, which left that local unused. This switches to column.flatten(), which respects the list array's own offset, per column so each keeps its own offsets.

Only wrapped is affected. bfd and bfd_split hit the same pattern but are protected by the pc.filter and pc.take that run before it.

Output is byte-identical to current behavior when the table is not sliced, so no existing correct result changes. Also checked on large_list with int64 offsets.

Tests

Added TestPackDatasetWrapped::test_with_multiple_map_batches, which packs across two map batches. It fails on main with the duplication above and passes here. The two existing wrapped tests use 3-row datasets, so they only ever exercise a single batch, which is why this was not caught.

pytest tests/test_data_utils.py -k "Pack"
10 passed, 548 deselected

Beyond the added test, I checked _pack_wrapped against an independent reference (concatenate each column's rows in order, then re-chunk) over 600 randomized shapes plus targeted edge cases: slices at every offset, seq_length of 1 and
equal to and greater than the total, ragged and empty rows, zero-row tables, multi-chunk tables, 1 to 3 columns, list and large_list, int32 and int64 offsets. All pass with the fix and 127 of them fail without it. pack_dataset also round-trips end to end across batch sizes 1 to 777, num_proc 2 and 4, and all three strategies, with input_ids / attention_mask / labels staying aligned.

Since #5189 was a performance PR, I benchmarked this too. flatten() is metadata work like values, so throughput is unchanged (at 500k rows x 16 tokens the median of 7 runs is 0.08 ms before and 0.09 ms after, which is within noise).

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

Who can review?

@qgallouedec @albertvillanova