feature: add asynchronous cycle which can deal with long-running evaluations

[hollandjg]

... and if everything crashes, doesn't completely lose its data.

It could be:

• a message queue
• a lightweight program which can run on any machine
• a lightweight program which runs as a cron-job

It needs to be able to:

• dispatch jobs to OSCAR (e.g. theorists, experimentalists), e.g,. using https://github.com/PySlurm/pyslurm
• dispatch experiments to the experiment runner,
• not hog expensive resources like GPUs if it doesn't need them right then

[hollandjg]

Our current approach relies on the Controller and Cycle objects, and running within python. We've demonstrated saving and loading (though it's not yet in a PR which is ready to go), as well as being able to handle many tasks like seeding and different processes depending on the current state within the framework of the Controller.

⛔️ However, we still don't have robust SLURM dispatch, which is a standard problem for data processing pipelines. It turns out that this is hard.

We should switch gears and look at pipeline tools. Can we set up the cycle to run under:

• cylc?
• snakemake?
• make?

[hollandjg]

We have a working version of the workflow module which does this. Open:

• Thorough testing
• more documentation

[hollandjg]

This is implemented with autora-workflow version 0.5.0 with the State mechanism. Needs testing in the real world.

[hollandjg]

This is now possible using cylc and the workflow package. Hurrah!