feat(tokenization): Added flag to run tokenization on multiple cores

[talolard]

I added a flag that runs the tokenization step on multiple cores per #57.

Ended up doing this with worker processes that read and write to a queue, so that each process loads spacy only once.

Seems to be 2x faster with 3 cores. I tested

• 50K documents
• 146,813,409 "words" (per wc)
• Using spacy tokenizer
• On 4 cores (so 3 in use because we use n-1 cores)
• times are for the entire process with count vectorizer
  With multiproc
  
  Without multiproc
  

Possible improvements

• Figuring out a way to drain the output queue faster would help this scale to more cores.
• Reading the file in the children processes instead of the parent would reduce IPC. Not much of a gain though because we're constrained on clearing the output queue

[talolard]

Don't merge this yet I found a bug