add multi30k for eval

[elliottd]

Add the Multi30K datasets for multilingual image--sentence retrieval evaluation. The evaluation data is available in English, French, Czech, and German. The sentence data can be found on Github at https://github.com/multi30k/dataset/tree/master/data/task1/raw.

The raw untokenized sentence data can be found in the following files, where LANG = (en, cs, de, fr):

test_2016_flickr.LANG.gz
test_2017_flickr.LANG.gz
test_2018_flickr.LANG.gz

The corresponding image information can be found in https://github.com/multi30k/dataset/tree/master/data/task1/image_splits

test_2016_flickr.txt this uses the test set images from the original Flickr30K dataset.
test_2017_flickr.txt uses newly collected images
test_2018_flickr.txt uses newly collected images

The newly collected images are available to download via Google Drive. Not sure if this is easy to automatically download so re-hosting elsewhere might be possible.

[rom1504]

seems good

we would need something similar to https://github.com/LAION-AI/CLIP_benchmark/blob/main/clip_benchmark/datasets/multilingual_mscoco.py

autodownloading gdrive is possible, done for a few datasets here already