Add num_processes to reader.train() to configure multiprocessing

[tholor]

In some cases (e.g. debugging or fine-tuning on small datasets) it can be useful to disable multiprocessing.
Let's add an argument to reader.train() that allows this.

Usage:

num_processes = 0 or 1  --> Plain, single python process 
num_processes = 12 --> Multiprocessing with 12 processes
num_processes = None --> Multiprocessing with "number of CPU cores -1" processes

Related to #268

[Timoeller]

looking good.

I changed the PR message, since when max_processes is either 0 or 1 no multiprocessing pool is initialized.

I also made use of num_processes in the FARM Inferencer and max_processes in FARM DataSilo more consistent in deepset-ai/FARM#480