Replace CroissantML with custom cache management

[MarcT0K]

Summary

Replace CroissantML-based data handling with a custom caching and dataset management system.

Changes

• Removed dependency on CroissantML
• Introduced a custom cache layer for dataset download and storage
• Added automatic update detection using dataset metadata (dateModified)
• Implemented cache_only mode for offline/reproducible usage
• Added streaming downloads with progress (requests, tqdm)
• Ensured atomic writes to avoid corrupted downloads
• Normalized dataset structure (full/ and reduced/ directories)

Impact

• Removes external dependency (CroissantML)
• Improves control over caching and updates
• Enables offline usage via cache_only

Notes

No breaking changes to the public graph loading API.