Fix FairseqHydraPretrainJob for better start_checkpoint behavior

[AndreasPlt]

Currently, when fairseq_hydra_config["checkpoint"]["restore_file"] is given as a parameter for the pretraining, this makes resuming jobs much more difficult since it will always restart from the given file - even though the model might have been trained for several more epochs already. Since changing the restore_file parameter to the new checkpoint would change the job parameter again, the job would also not be able continue training again.

I therefore propose to handle the "restore_file" parameter in the job individually and not leaving it to fairseq. The idea is to save the given "restore_file" as a job attribute to later move it to "output/checkpoints/checkpoint_last.pt" if the latter does not exist yet, which is also the default parameter inside fairseq. That way, when training has been already done for several epochs, the job can still continue from checkpoint_last.pt, and use the given checkpoint if the training hasn't already ran earlier.