[feat] Option to customize converter instructions for pydantic output

[eelkevdbos]

Currently, the instruction to convert raw llm output to structured pydantic models is hardcoded to: "I'm gonna convert this raw text into valid JSON.". I ran into some cases where this resulted in incorrect transformations with "too much" JSON.

For example with this simple pydantic model:

...
class SimpleResult(BaseModel):
  result: str

Often times, when the llm would output a simple scalar value, it would start to improvise leading to something resembling:

SimpleResult(result='{"result": "1"}')