Make applying embeddings more efficient.

Adding new tokens no longer makes a whole copy of the embeddings weight
which can be massive on certain models.