`Simulation`'s `data` input option accepts arbitrary dictionaries

[anth-volk]

Based on the documentation here, The Simulation initializer's data arg accepts an arbitrary dictionary as an input, but will only work if this dictionary is a properly-configured household object. This dictionary (it appears) is only used if the simulation type is household-level, while a macro-typed sim expects a str | policyengine_core.Dataset | None-type.

This introduces quite a few drawbacks:

1. If data can be str | Dataset | dict[Any, Any] | None, that's basically every type; we start to lose the advantages of applying typing
2. data must be handled by both the household and macro calculation processes, meaning both will need to evaluate the existence of any of these types
3. Passing an arbitrary dict will mean we either need manual validation on the household calculation side, or we will create strange breakage patterns for users unacquainted with the schema
4. The datasets we use are, under the hood, merely an array of households. Let's say we allowed the passage of custom subsets of households for calculation purposes. Because we've tied ourselves to a dict, this would require schema migration. I think it'd be better if we just started out using a list[dict] structure and just confirmed for the time being that these are just single-entry lists.

[anth-volk]

Another point on this: household-level calculations accept pretty different setup args than macro. Why not have a households: list[Household] arg (with Household defined to parse and validate the household schema as it currently stands), then a separate dataset: Dataset | str | None with the behavior we currently have? You could even break that out into dataset: Dataset and dataset_url: str | None and just prevent issues by failing if both are provided.