Fix inconsistency in fluiddyn.cluster

Here, I wrote:

I think there is an inconsistency between different flavors of clusters in fluiddyn.clusters. It has to be fixed.

For SLURM and PBS clusters, if nb_mpi_processes is not set (leading to nb_mpi_processes = None), the default value is nb_cores_per_node * nb_nodes. It is not the case for OAR clusters.

We actually already use this behavior at LEGI to launch fluidimage processing without MPI.

I think we anyway need to "break the world" for at least some clusters. I tend to think the behavior we have for OAR is more reasonable and better (for fluidimage, snek5000, dask, etc.).

We could support nb_mpi_processes="auto" to specify that we want to launch a MPI command without explicitly computing the number nb_mpi_processes.

This way, it would be much simpler to launch snek5000 simulations on Slurm and PBS clusters.

@avmo would it be alright for you? I guess you would need to adapt some of your scripts...