Some fixes, get rid of data_task, homogenize model_task to model_type

[francoishernandez]

This PR regroups a few fixes/enhancements:

• ~data_task is deduced from encoder/decoder presence in model rather than edgy loop over data;~ -> data_task is no more, using model_type instead;
• restrict share_embeddingsto EncoderDecoder/seq2seq setup (enabling it by mistake on some Decoder/LM run prevents proper saving of tgt_emb...);
• samples are saved in save_data rather than its parent dir;
• move update_config_with_checkpoint prior to prepare_transforms_vocabs in training initialisation, to properly support update_vocab.