Initialize pretrained word embeddings

Eole prepares pretrained embeddings during eole train initialization. Supply the embedding text file and type in a complete training YAML:

save_data: run/embeddings
both_embeddings: glove_dir/glove.6B.100d.txt
embeddings_type: GloVe
model:
  architecture: transformer
  embeddings:
    word_vec_size: 100
    freeze_word_vecs_enc: false
    freeze_word_vecs_dec: false

Use src_embeddings and tgt_embeddings for separate files. GloVe and word2vec text formats are supported. Embedding dimensions must match the model; set the freeze fields under model.embeddings to keep those weights fixed. Matched tensors are saved as <save_data>.enc_embeddings.pt and <save_data>.dec_embeddings.pt. Download and extract the selected vectors before training; the placeholder path above is not a bundled asset.