Support the Open Images dataset

[IRDonch]

Open Images is a dataset with ~9 million images and annotations of various types:

• image-level labels;
• object bounding boxes;
• instance segmentation masks;
• object relationships;
• localized narratives (synchronized audio recordings, text transcriptions and mouse traces describing the image).

The first three types are directly supported by Datumaro. The last one is not fully supported, but could be used to create caption-type annotations.

I think it would be reasonable to limit the scope of this issue to annotation types that are already supported. The ones that aren't can be added in the future.

[IRDonch]

Hmm, looks like GitHub was overeager in closing this issue. #291 only partly resolves it. :-)