[WIP] DINO self-supervised recipe with VoxCeleb

[JaejinCho]

This is a work in progress (WIP). Started from the persephone branch's latest commit while closing the old one: #59.

DONE

• Working scripts for training w/ single GPU (only runnable in the JHU CLSP grid using already prepared data by following list_run.sh)

TODO (maybe in order)

• Multiple-GPUs
• This actually worked fine in the older version of hyperion so the error might be due to version mismatches in some packages.
• Data Preparation

[JaejinCho]

@jesus-villalba

• Please let me know if this PR is better to be merged to a new branch having the same name as my fork's branch (persephone_dinossl)
• I mentioned "The code for experiments will be uploaded to the public repository
  https://github.com/hyperion-ml/hyperion." in our Interspeech paper, Non-Contrastive Self-Supervised Learning of Utterance-Level Speech Representations. I though mentioning this branch in master's Readme can help people to find this code once finished.

[JaejinCho]

@jesus-villalba I've just fixed the rest check boxes above and removed some unused configuration files. You can check this PR now. Meanwhile, I may check the overall scripts checking for this PR lastly.

BTW, the evaluation script is not been written yet after training, but I believe this is straightforward (mainly getting a trained teacher model to extract embeddings). If I could add it, I would do so in another PR. Let me know if you have any comments.