Add MetadataAnnotations class for user-facing handling of annotations

[kbolashev]

Scope of this PR:

• Add a new dependency: https://github.com/DagsHub/dagshub-annotation-converter/
• Plus additionally add ultralytics into dev dependencies
• Add MetadataAnnotations class. Objects of this class will get created for any annotation field in the QueryResult, so you can use them like this:

dp = ds.all()[0]
anns: MetadataAnnotations = dp["exported_annotations"]
anns.add_image_bbox("person", 0.1, 0.1, 0.1, 0.1)
anns.add_image_bbox("cat", 0.2, 0.2, 0.1, 0.1)
dp.save()

• Autoload annotations on ds.all() and ds.head()
• Add an ability to turn off this autoload functionality as load_documents and load_annotations kwargs in these functions

[dagshub]

Join the discussion on DagsHub!

[simonlsk]

All very cool features, added some comments and questions.
Thanks!

[simonlsk]

• Remember

[simonlsk]

This bit is unclear to me. You discard load_into_memory and force it to be true. Is there a common scenario where you pass load_into_memory=False explicitly to get_annotations, and expect this to happen?

[kbolashev]

No, however I pass the kwargs to the downloading functions, and loading annotations requires them to be already in memory, so load_into_memory=False will break things.
Actually I should probably handle this particular edge case better in get_blob_fields and either skip loading if load_into_memory=False (explicitly respecting the user, even if it doesn't make sense), or force-opening.