External embeddings in multilingual parsing

[mdelhoneux]

We currently load external embeddings on a per-language basis and overwrite any word that was seen in a previous language. Would be good to do something smarter.