Merge branch 'main' of https://gitlab.version.fz-juelich.de/MLDL_FZJ/juhaicu/jsc_public/sharedspace/teaching/intro_scalable_deep_learning/course-material into main

When connecting with X11 forwarding, the `DISPLAY` variable is set and
passed to Slurm jobs. This causes plotting to error out. We now
explicitly select a non-interactive matplotlib backend.

And whitespace

These don't require a PDF re-build.

c68a84a4 introduced some errors.

We now handle the missing batch size argument so these possible mistakes
are more obvious.

The APIs haven't been experimental for a long time.

The API hasn't been experimental for a long time.

This just causes confusion and a bunch of redundant output if people use
it with the module system.

More Pandoc/LaTeX build files.

Merge branch 'main' of https://gitlab.jsc.fz-juelich.de/MLDL_FZJ/juhaicu/jsc_public/sharedspace/teaching/intro_scalable_deep_learning/course-material into main
merge before push

`lr` argument is deprecated, use the new recommended name
`learning_rate`.

And related Python versions.

This way, participants can try out Horovod commands interactively.

The previous explanation referred to how `tf.data.Dataset.shard` shards
data; in practice, this doesn't make a difference and is simply
confusing to readers.

Instead of `tasks-per-node` which is not documented.