Self-trained fasttext embeddings

[michelole]

To better compare with pre-trained embeddings.

Draft idea:

• [x] Dump n2c2 train data into a format legible by fasttext
• [x] Harmonize pre-processing (#105 and #86)
• [x] Train fasttext embeddings using command line
• [x] Extract subword embeddings for train/test data
• [x] Retrain LSTM using such embeddings

[michelole]

fasttext training follows https://github.com/ncbi-nlp/BioSentVec/blob/master/src/train_biowordvec.sh

[michelole]

Reopening while models don't finish training.