Finetuning bugs

[saattrupdan]

This fixes several issues with the finetuning:

1. Interleaving processed datasets sometimes caused issues, which was fixed by applying dataset processing after the interleaving.
2. When finetuning Whisper models, use the WhisperConfig.max_length as an upper bound for the model_max_length of the tokenizer. Previously we just had 512 as the upper bound, but some Whisper models require shorter context lengths. This context length is for transcriptions (i.e., labels) only.
3. Previously in multi-gpu setups, we had to set padding to 'max-length'. I don't recall why this was the case, but this now (for some reason) leads to issues. I've commented out the "forcing it to max-length" block now, along with a TODO comment that this might change in the future.
4. PyTorch dataloaders seem to try to be clever and splits up the batches for each device in a multi-gpu setup. But since we're already handling the splitting, we need to set dispatch_batches=False.