Motivation:
1) Now we have to predict how much RAM we should reserve for single jobs. If we choose 1G, then we are lowering our possible concurrency. If we choose 512M, then worker could potentially lack memory.
2) We have different configs for judging machines which vary only in concurrency.

What to do:
1) Replace concurrency in worker config with RAM capacity
2) OIOIOI should add to environ information how much RAM single job needs
3) Scheduler should operate on RAM capacities instead of concurrencies.
4) Split this task into smaller ones :)

Adnotation:
Don't forget about taking into account number of CPU cores. Currently used formula is min(# cpu cores * 1.5, (RAM minus a little bit for OS) / (max required mem per problem)). Obviously it is worth to discuss if and how this formula should be changed.