Train eynollah on labeled data

[kimlee87]

• Update function labelstudio_to_png() in groundtruth.py
  • to draw rectangles at correct positions (x and y from LabelStudio is the top left corner, not the center)
  • add suffix _dup to 18 duplicate image names with an overwrite flag so we can use 82 or 100 annotated images (overwrite=True or False, respectively)
• Preppare data for training by grouping by year and page type before shuffling and splitting
• Document setup steps for training Eynollah
• Add a script to download original files corresponding to labeled files, as well as unseen data from SDS.