Enabling DeepSMILE on large encoded datasets

[vale-salvatelli]

This PR contains two changes necessary to run DeepSMILE on a large dataset when using the Innereye SSL checkpoint (or any other encoder with high dimension):

• option to encode in chunks (this prevents OOM error when performing the encoding)
• option to load the cached encoded dataset in CPU (this prevents OOM when loading from the cache)

It also changes how the PNG images are loaded all over the histo pipeline to make the loading faster (see https://hi-ml.readthedocs.io/en/latest/loading_images.html)

Please follow the guidelines for PRs contained here. Checklist:

• [ x] Ensure that your PR is small, and implements one change.
• [ x] Add unit tests for all functions that you introduced or modified.
• [ x] Run PyCharm's code cleanup tools on your Python files.
• Link the correct GitHub issue for tracking.
• [x ] Update the Changelog file: Describe your change in terms of
  Added/Changed/Removed/... in the "Upcoming" section.
• When merging your PR, replace the default merge message with a description of your PR,
  and if needed a motivation why that change was required.

[ant0nsc]

Looking good, just a few minor comments