Intrgrate Didem's visualizing image embeddings

[Ultimate-Storm]

Regarding the visualization pipeline you are developing and gathering, Didem has this repo on the groups' github called visualizing image embeddings: https://github.com/KatherLab/visualizing_image_embeddings. It takes images/tiles and labels (for the purpose of coloring plots according to class). Upon reducing dimensions with t-sne, it plots both a plain scatter plot with dots and a scatter plot with the images on the place of images instead of dots.

• add proper readme for this project
• made it a proper python package
• format input arguments
• add example output images
• store model ckpt elsewhere
• test on Marko's cohort and evaluate the results

[dcifci]

What can be improved in visualizing_feature_embeddings:

• implement RetCCL feature extraction, remove the existing SSL model
• Feature extraction must be done on batches
• Feature extraction code is unnecessarily modular, which makes things more complicated and harder to debug (sorry about that!)
• Two plots are plotted with different libraries: make them compatible
• In the thumbnail plotting, image size must be adjustable by the user so that one should not go over the source code every time. We can't decide on thumbnail size automatically. The solution I can recommend is that feature extraction and clustering/plotting must be two different steps.
• give more more user-defined options for plotting (advanced task:))

[dcifci]

Sample input and output added, and readme is updated. Readme could be more detailed, but it makes more sense to do it here in wanshi utils.

[dcifci]

• correct dot plot legends for showing actual names

[dcifci]

• create a unique name for each run, maybe with user input