Potential better structure for downloaded transcriptions zip file.

[harrykeightley]

Currently, for inference performed with hugging face models (not local pretrained models), some workarounds were required because the model names have forwardslashes in their name e.g. facebook/wav2vec2-base....

To get around this, we base64 encode the model names on the server to give them standardized, unique paths for when we try and make a folder that links with the model name.

When downloading the zipped transcription files, users might prefer to see the original model name (somehow) than the b64 encoded versions.

This is not a high priority issue, just recording it here so that I don't forget.