Fix/smaller fixes

[saattrupdan]

This PR does the following:

1. Replaces isort and flake8 pre-commit plugins with ruff, which is faster and has more features.
2. Specifies the common voice version in the dataset, to allow better reproducibility.
3. Make it possible to train on local datasets, since using streaming with large datasets seem to be quite unstable, as Hugging Face's servers seem to crash all the time.
4. Add ctc_zero_infinity=True, which might fix the NaN loss issue (see Replicate NST fine-tuning on Alvenir machine to sniff out training or data issues #32).
5. Add first/second momentum of Adam to the config.
6. Use the new-ish auto_find_batch_size=True argument.
7. Fixes an issue which caused bugs when trying to build NST-da in a directory that isn't the current working directory.
8. Don't include NST-da samples with empty transcriptions anymore.
9. Change the no_cuda argument in TrainingArguments to the newer equivalent use_cpu argument.

[sorenmulli]

Looks good, two small comments