Use a different experiment folder's checkpoints as the base model when running an silnlp experiment on clearml

[laura-burdick-sil]

Enable multi-stage fine-tuning of NLLB, by setting the "model" parameter in the config file to a previously trained model, rather than the base huggingface model.

Damien (slack): "Our old OpenNMT implementation supported parent models and multi-stage finetuning. I guess we never implemented it for Huggingface. You can even see that there is a has_parent property on the Config class that isn't used for anything after we removed OpenNMT support. You would configure it by setting the data/parent setting to the experiment name. Huggingface transformers definitely supports fine tuning from a local checkpoint, so we should be able to implement it."