Implement flexible parallelization

[RandomDefaultUser]

For large scale simulations, we need to have MALA highly-parallel. More precisely, we need to address the following:

• [x] Parallel descriptor calculation
• [x] Parallel network passes
• [x] Parallel QE interface
• [ ] Clever resource handling; What I am thinking of is an automated workflow of this sort, that is able to handle an inhomogeneous amount of ranks, i.e. using a lot of CPUs for the CPU heavy things (SNAP descriptors, QE), but fewer GPUs for the NN passes; in the best case with some kind of load manager for e.g. Monte Carlo simulations

The first two have been adressed already. For the latter, I envision something like he following:

[RandomDefaultUser]

I think the current version of parallelization (except the GPU stuff, which we are working on) works well enough for now. Closing this, since we want to shift inference to GPU anyhow.