fix(datasets): use len(lengths) for default modes in multimodal add_document - #5726

What

IndexedDatasetBuilder.add_document documents modes as defaulting to None. On a multimodal builder (multimodal=True), the default path built the per-item mode list as [0] * lengths:

if self.multimodal:
    self.sequence_modes.extend(modes if modes is not None else [0] * lengths)

lengths is a List[int], so [0] * lengths multiplies a list by a list and raises TypeError: can't multiply sequence by non-int of type 'list' — it never produces the intended default mode list. Any multimodal dataset build that relies on the documented modes=None default hits this.

No linked issue — found by reading the code.

Fix

Use [0] * len(lengths), so the default assigns one mode (0) per item — matching the one-mode-per-item contract that add_item and sequence_lengths already follow.

Test

Adds test_multimodal_builder_add_document_default_modes in tests/unit_tests/data/test_builder.py: a multimodal builder calling add_document with the default modes no longer crashes, assigns one default mode per item (matching sequence_lengths), and still honors an explicit modes list. It fails before the fix (TypeError) and passes after.

black (24) + isort + ruff are clean on the changed files. The full unit suite runs on GPU CI; I verified the fix and the add_document body standalone off-GPU (the old expression raises TypeError, the new one yields one 0 per item).