Revise the MNIST example, cleaning up the code and documentation.

The MNIST example, in its non-quickrun flavor, has some documentation/comments artifacts from being derived from the Heart Disease UCI one, and relies on declearn-split, which does not feel entirely proper to me as it makes up for an in-between using plain scripts and CLI entry-points.

This small MR remedies to the former issues, and also adds data-preparation to the "main" run_demo.py script, which makes up for a one-liner demonstration of DecLearn.