fix CI for datasets v4.7.0

[snimu]

Description

Starting with v4.7.0, HuggingFace datasets serialize heterogeneous content-part dicts (e.g. text + image_url) to JSON strings when storing in Arrow tables. Recover these by attempting json.loads on string content parts during message normalization.

This is a quick and dirty fix for now. #1002 could be an opportunity for a more general fix.

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

---

Note

Low Risk
Small, localized normalization change with best-effort JSON parsing; main risk is subtle behavior changes if arbitrary string content happens to be valid JSON.

Overview
Fixes message normalization for HuggingFace datasets v4.7+ where heterogeneous content part dicts may be persisted as JSON strings.

During _normalize_raw_message_content, string parts are now best-effort json.loads-decoded and only dict parts with a type field are converted via from_raw_content_part, improving robustness when reading dataset-backed messages.

^{Written by Cursor Bugbot for commit 5817f8c. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

json.loads silently converts strings to non-dict types

Medium Severity

json.loads on a string content part can return non-dict types (e.g., "null" → None, "123" → 123, "true" → True, "[1,2]" → list). These silently replace the original string in part, bypass the isinstance(part, dict) guard, and get appended as invalid ContentPart types. The parsed result is only useful when it's a dict, so the assignment to part needs to be guarded — e.g., only replacing part when isinstance(parsed, dict).