refactor: SummarizationAccuracyMetrics transform to handle multiple target outputs more efficiently

[kirupang-code]

• Edited the logic in SummarizationAccuracyMetrics so that the target_output field wouldn't have to be duplicated in order to be evaluated for each model_output. If we have multiple target_outputs we take the max score from self.compute_metric.
• Added an additional target_output_keys_provider parameter to support a way to compute scores given a key to a list of possible target outputs.
• Updated the ordering of subclass transforms in order to keep everything consistent with their parent transforms, otherwise parameters would get mixed up. I had to add a default value for BertScore because it follows a parameter with a default value. input_keys + model_output_keys, TypeError: can only concatenate str (not "list") to str

[danielezhu]

Let's keep this as a positional argument so that 1) the argument order doesn't change, which would break things for existing users, and 2) the default configuration isn't invalid; currently, if you use the default options, you'll get a validation error.

For point 1), I really mean that we don't want the signatures of the concrete subclasses to change. Obviously, no one is directly calling this ABC's init method.

[danielezhu]

We need to put target_output_keys_provider as the last argument, or else existing calls to this method will break.

[danielezhu]

Discussed with @kirupang-code offline. We need to use this ordering (which is backwards-incompatible) in order for deserialization of these object instances to work properly (since we rely on a specific ordering of the arguments to __init__). The existing calls to BertScore.__init__ in our own code have been updated as necessary, but users who interact with BertScore objects directly will be impacted.

We should keep track of this before the next release.

[danielezhu]

Why is this needed?

[danielezhu]

Discussed with @kirupang-code offline. We need to use this ordering (which is backwards-incompatible) in order for deserialization of these object instances to work properly (since we rely on a specific ordering of the arguments to __init__). The existing calls to BertScore.__init__ in our own code have been updated as necessary, but users who interact with BertScore objects directly will be impacted.

We should keep track of this before the next release.

[danielezhu]

Please add a unit test where there are multiple target outputs, and verify that the max score is returned.

[kirupang-code]

Had to add # type: ignore to target_output_keys_provider because of its Optional type since we validate that its value is not None. Discussed with Daniel. Errors without the # type: ignore:

3. Mypy
=======
src/fmeval/transforms/summarization_accuracy_metrics.py:89:69: error: List item 0 has incompatible type "str | None"; expected "str"  [list-item]
src/fmeval/transforms/summarization_accuracy_metrics.py:111:25: error: Invalid index type "str | None" for "dict[str, Any]"; expected type "str"  [index]

[xiaoyi-cheng]

lgtm