refactor evaluator base and task classes

[lvwerra]

Refactored the evaluator base and task classes based on discussions in #178 to make them a bit more abstract for new tasks and draw a clearer separation between base and task classes.

Adding the performance should now be easy: measure inside call_pipeline and return an dict with which you can update the result dict inside the compute methods.

Also if a user wants to pass a custom prediction postprocessing (as suggested by @craffel) before passing to the metric we can just overwrite the predictions_processing method or wrap the custom function around it.

Let me know what you think @ola13 and @fxmarty.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[ola13]

Hey, thanks for this! I posted a comment here #178 (comment) - can we discuss there for now since there's a bit more context there? :)

[lvwerra]

Thanks for your feedback @ola13. I integrated the suggestion and think I found a way to keep the base class abstract enough that in most cases just the prediction_processor needs to be adapted.

I also added a PIPELINE_KWARGS and METRIC_KWARGS attribute to the class such that we can control the kwargs especially for the pipeline (truncation=True). That way if a user really wants they can adapt the runtime settings of those to classes. This also addresses #137.

Let me know what you think.

[fxmarty]

Looks fine to me!

[ola13]

LGTM 🚀