Provide clearer documentation for data download operations

[anth-volk]

At the moment, the data download operations we use are unclear, poorly documented, and concerningly typed (e.g., dataset can be a string or a Dataset type or None, and we make a lot of in-place modifications that change types without making this clear). We also appear to try to support all types of interfaces and implementations, but make that unclear; e.g., it seems I could initialize a simulation and override our dataset option to (attempt to) download literally any Hugging Face dataset by prefixing it with hf://, though this will fail because the dataset must exist within our Hugging Face repo.