Investigate multiprocessing memory usage further

[bengioe]

It's not clear that we are doing multiprocessing in the best way possible.

• we have been seeing problems with too many shared tensors being created, the current patch to that is Reduce load on shared memory #59 (which naively pickles everything!)
• we obviously waste time copying tensors and objects from/to workers

A great enhancement for this project would be to dig deeper into those problems.