Implement a `predictions_to_coco` function

[niksirbi]

We currently have annotations_to_coco, but we also need to convert prediction files to the COCO .json format, to facilitate the extraction of cliplabels.json (see #10).

For SLEAP .slp file, it could be that annotation_to_coco can already do this? Though we'd need to double-check how labeled vs predicted instances are handled by sleap-io here.

For DLC (and LP) files, I suspect we'll for sure need a separate predictions_to_coco function. We'd discussed maybe doing it via movement (e.g. DLC predictions -> movement dataset -> SLEAP file -> sleap-io -> COCO), but:

• Not sure that would work
• Probably it's more efficient to not go via movement in this case.

Note

Because we are talking about predictions inputs here, we can probably use movement sample data for testing.