Update DocumentLSTMEmbeddings

[tabergma]

Currently, you can set parameters bidirectional and use_first_representation independently. As it does not make sense to train a LSTM bidirectional and just take the last representation, we should concatenate the last and first representation as soon as the parameter bidirectional is set. Thus, we could remove the parameter use_first_representation.

Additionally, we need to register the token embeddings as module to avoid cuda - cpu conflicts.

for i, embedding in enumerate(embeddings):
    self.add_module('token_embedding_{}'.format(i), embedding)