[feat] support messages column as JSON string in iterable datasets

[mwxely]

Motivation

Parquet's Arrow schema inference struggles with deeply nested, heterogeneous structures like OpenAI-format chat messages. A single messages column may contain varying content types (text, image_url, video) with different nested fields across rows, leading to schema conflicts during both file creation and multi-file loading.

A pragmatic and widely-used workaround is to store the messages column as a JSON-encoded string. This is already common practice in the community — many ShareGPT-format datasets on HuggingFace Hub use this approach. However, the current iterable dataset loaders assume messages is always a native list[dict] and crash on string input with:

TypeError: string indices must be integers, not 'str'

Modifications

Add an isinstance(messages, str) check with json.loads() fallback at all entry points where messages is read from data:

• Qwen3VLIterableDataset.load_from_json() — handles video/image multimodal data
• VisionSFTIterableDataset.load_from_json() — handles image-only data
• VisionSFTIterableDataset.load_from_hf() — handles HuggingFace dataset format

The change is backward-compatible: native list[dict] messages pass through the isinstance check with zero overhead.