fix: Decimal in structured generation leads to errors

[andreatgretel]

During structured generation, quantities like prices and such are typically modeled using Decimal, see e.g. our tutorial number 2.
Some models will output such quantities both with and without quotes. We currently use gsonschema within the StructuredResponseRecipe to validate these outputs, and for Decimals, Pydantic gives for the schema:

"anyOf": [
  {"type": "number", "minimum": 10.0, "maximum": 1000.0},
  {"type": "string", "pattern": "^(?!^[-+.]*$)[+-]?0*\\d*\\.?\\d{0,2}0*$"}
]

Thus, both strings and floats are accepted, which will cause an issue when calling dataframe.to_parquet. This has been happening often in our tutorial number 2.

The proposed solution is to convert fields detected as Decimals to string, always. This is done inside gsonschema.validators.validate. A small test is added to ensure the behavior.

[johnnygreco]

Is this for sure the direction it needs to go? i was thinking that the string would need to be converted to a float. could definitely be wrong, though

[andreatgretel]

I think the point of Decimal is being able to control the number of digits after the decimal point. That is better accomplished by a string I believe? Floats can approximate it incorrectly etc.

[andreatgretel]

Alright, following offline discussion - I'm first converting it to Decimal to ensure the right precision, then finally to float as the final format.