Pinning memory can cause errors for smaller GPUs

[jorshi]

I'm getting a memory error trying to run predictions: RuntimeError: Caught RuntimeError in pin memory thread for device 0.

Setting pin_memory=False in the DataLoader in task_predictions.py solves this issue. https://github.com/neuralaudio/hear-eval-kit/blob/174b9268f96d3b4de51cd3557f9063d64fe1bd02/heareval/predictions/task_predictions.py#L744

Is this just for speed? Can we make this optional?

[turian]

Yes, we can definitely make this option. We might consider making a config.py, separated into configs that affect results (this one doesn't) and ones that are basically just performance