This PR adds word embedding (Glove, Word2Vec, Fasttext) based Language Models to FARM - in a very basic way

General idea: Instead of contextualizing word embeddings through BERT we just convert words to its corresponding word embedding. The Prediction Head on top of the Language Model now does all the heavy lifting.

Example scripts for:

• using Glove or Word2Vec as base Language Model for doc classification
• using Fasttext as base Language Model for doc classification
• embedding extraction

Features:

• get pretrained German Glove and fasttext Embeddings from s3 through fasttext-german-uncased and glove-german-uncased
• conversion script from gensim Glove/Word2Vec models to FARM models

Limitations

• the embeddings are static. They do not change over the course of training
• no saving or loading of finetuned models (on a downstream task)
• prediciton heads only support Feed Forward Neural Networks
• no multiprocessing for data conversion
• performance on tasks like GermanEval18 is inferior to Bert based approaches