Update the old tutorials

[jarlsondre]

The old tutorials are quite messy and outdated. They still use the old SLURM scripts and contain a ton of code that isn't necessary, confusing the user. They should be updated to match the current state of the library.

Tests known to be outdated:

• "1. Introduction to distributed ML with PyTorch"

[okrochak]

When running torch-tutorial-1-mnist, horovod has 4x as many examples, so 4x longer training when compared to deepspeed or ddp.