Use multiprocessing.shared_memory to reduce pickling overhead in MultiprocessingEvaluator

[EwoutH]

I noticed in a profile run of the MultiprocessingEvaluator that pickling takes up a signficant amount of runtime. Since we are using Python 3.8 or later, we can leverage the multiprocessing.shared_memory module to share large data structures between processes without having to pickle them. This can significantly reduce the serialization overhead. We can create a SharedMemory object to store your data, and then use ShareableList or numpy arrays to interact with the shared memory.