Composition and sematic predictions commented out

[dveni]

Those features are currently commented out (https://github.com/mediatechnologycenter/aestheval/blob/main/aestheval/data/reddit/prepare_dataset.py#L296) because the merge is on the im_paths (which changed), shouldn't be hard to fix but I have to find time to write a better implementation of this.