Add SODA-A dataset

[nilsleh]

This PR adds the SODA-A dataset. Dataset rehosted on HF.

Dataset features:

* 2513 images
* 872,069 annotations with oriented bounding boxes
* 9 classes

Dataset format:

* Images are three channel .jpg files.
* Annotations are in json files

TODOS:

• some annotations have more than just 8 coordinates, for example file Annotations/train/01874.json:

{
        {
            "poly": [
                1248.874059994201,
                2785.0,
                1267.2565270181412,
                2785.0,
                1290.0001220703125,
                2783.999755859375,
                1289.2877197265625,
                2767.801025390625,
                1248.1971435546875,
                2769.608154296875
            ],
            "area": 659.4863047840048,
            "category_id": 7,
            "image_id": 1874,
            "id": 43
        },

Example plot:

[nilsleh]

@shaunyuan22 Thank you for this nice dataset and all the work. We aim to make the dataset more easily usable in torchgeo, and would appreciate it if you have any comments, corrections etc.

[nilsleh]

Open question how to deal with the polygons into a common oriented bounding box schema.

[adamjstewart]

Asked about oriented bboxes on Slack. We may need to add support to Kornia for this ourselves. Until Kornia supports it natively, I guess it doesn't matter what the format looks like. But let's use the same key names that Kornia uses.

[nilsleh]

Instead of open-cv, as an additional dependency, I implemented the bounding box logic with shapely:

[adamjstewart]

Most of my remaining comments are optional pedantic things, this is 99% ready to merge.

[adamjstewart]

Might be more clear if you use einops.rearrange here