More comprehensive multiprocessing

[sienkie]

TODO:

• Add multiprocessing to "prepare" mode, to have every network layer prepared in parallel. Make sure it performs well memory-wise for both high number of layers and large feature matrices.
• Update multiprocessing in "run" mode. Currently we are running factorization in parallel for every sparsity value. If we run it on parallel for every (sparsity, repetition) pair, functions would be more atomic and we would make better use of resources.

[sienkie]

NOTE: "prepare" could also be made more memory efficient if we could delete every feature matrix from memory after the corresponding adjacency matrix was created