Reduce model load time with quantized embeddings

This release contains one large change: the loading of quantized models is speeded up by computing the unknown word embedding as an avarage of the subquantizers, rather than an average of all in-vocab word embeddings.