Attention weights from attention decoder

[eric-ayllon]

forward_decoder method does not provide an interface for deactivating the keep_all_weights option of the decoder forward, forcing the attention weights to use a lot of memory even if they are not needed.