Refined prompt construction for feedback

[piotrm0]

Items to add to release announcement:

• Heading: delete this list if this PR does not introduce any changes that need announcing.

Other details that are good to know but need not be announced:

• There should be something here at least.

Work in progress. Designing prompts for feedback from several common parts and allow different sizes of such prompts depending on the allowable space:

class ScoringPromptBase(SerialModel):
        """Common parts to build a scoring prompt out of."""

        prefix_template: str = """"""
        """Text to include before all other parts."""

        interp_template:str = """You are a {purpose} scorer."""
        details_template: str = """"""

        low_score: int = 1
        high_score: int = 10

        shots_template: str = """EXAMPLES:"""
        shot_template: str = """
INPUTS:
{shot_inputs}
EXPECTED OUTPUT: {shot_output}"""

        suffix_template = """
Answer only with an integer from {low_score} ({low_interp}) to {high_score} ({high_interp}).

INPUTS:
{inputs}
SCORE: """
        """Text to include after all other parts."""

    class ScoringPrompt(ScoringPromptBase):
        """Specific parts to build a scoring prompt out of."""

        interp: str
        """Minimal interpretation of the score."""

        low_interp: str
        """Interpretation of a low score."""

        high_interp: str
        """Interpretation of a high score."""

        details: str
        """Text to include after the purpose to provide some more details about
        the purpose."""

        shots: List[Tuple[Dict[str, str], int]] = []

        def build(
            self,
            max_tokens: int = sys.maxsize,
            inputs: Optional[Dict[str, str]] = None
        ) -> str:
            """Build a prompt for the given inputs while staying under the token
            limit.
            
            The built prompt will have at least:
                - filled prefix_template,
                - filled interp_template,
                - filled suffix_template.

            If space allows, it will also include:
                - filled details_template,
                - filled shots_template (if at least one of the below is
                  included),
                - filled in shot_template for each shot.

            Args:
                max_tokens: The maximum number of tokens to use. Defaults to
                    sys.maxsize (effectively infinite).

                inputs: The inputs to fill in the prompts
                    with other than the ones defining the scoring task. Those
                    are contained in self.
            
            Returns:
                str: The built prompt.

            Raises:
                ValueError: If the prompt would be too long to fit within the
                    token limit even at its miminum size.

            """