The overall goal of this PR is to standardize data formatting so that the outputs of one stage can be used as the input of another. And to avoid head aches.
Two new objects are instrumental in accomplishing this:

1. The Data Schema objects, which define column names in one place.
2. Dataset objects which handle both reading from input and writing to output.
   • The different types of datasets will read CSV rows and output the relevant object for each row e.g. TestItem for prompts, SUTInteraction for prompts-responses, and AnnotatedSUTInteraction for annotations.
   • The Annotation dataset serializes/deserializes annotations as json-strings. It can serialize pydantic objects or dictionaries. Annotations will always be deserialized as dictionaries.

Also in service of this goal, every output file (aside from metadata.json) is a CSV file in which one row corresponds to one prompt and one sut response and/or one annotation. No more annotations.jsonl.