[FEATURE] Investigate removing memory as a consumable resource

[cartalla]

• Investigate removing memory as a consumable resource and allocate 1 job per node.
  • This is more scalable according to https://slurm.schedmd.com/big_sys.html because the scheduler doesn't have to keep track of the cores and memory on each node.

Some experiments indicate that it might be more cost-effective to run 1 job per instance anyway rather than trying to pack multiple jobs per node.
The extra cost comes when an larger instance remains running by a small job so that it isn't idle.