Add code to generate mask images for training

Doc-UFCN needs specific mask images (one mask per image). The source code is already available in this project but it's dependant on Arkindex.

@NolanB already extracted the minimal source code for our training worker, so we should port this in a new module doc_ufcn.train.mask.

This module should have 3 methods:

1. generate_mask(image_width, image_height, label_polygons, label_colors, output_path)
   • label_polygons is a dict of list of polygons, where the key of each item is a label name
   • label_colors is a dict of RGB tuples
2. dependant resize_polygons
3. dependant split_polygons

# Example parameters
image_width = 500
image_height = 800
label_polygons = {
  "text_line": [
    [[0,0], [0, 100], ...
  ],
  "picture": [
    [[0,0], [0, 100], ...
  ],
}
label_colors = {
  "text_line": (255, 0, 0), # red
  "picture": (0, 0, 255), # blue
}
output_path = Path(path/to/file)

Rules:

• each label_polygons keys must have a matching label_colors
• check RGB structure (tuple or list of 3 integer items between 0 & 255)
• check polygons (tuple or list of list of 2 integer items, always positive)
• image width & height are positive integers