feat: Add BERT_SCORE to QAAccuracy and update unit/integration tests

[kirupang-code]

• Added BERT_SCORE to qa_accuracy.py by creating SplitWithDelimiter, a transform that uses a target_output_delimiter to split a target output string into a list of possible targets. This allows us to compute multiple BertScores and take the max over all possible targets.
• Additionally, the integration tests for qa_accuracy were taking over 10 minutes to run 100 records because of the runtime of the bertscore model, so I created a smaller dataset, triviaQA_sample_small with 4 records to be used.
• Updated qa_accuracy_semantic_robustness.py to import and use QA_ACCURACY_SCORE_NAMES (QAAccuracy metrics w/o BERT_SCORE) instead of SCORE_NAMES.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[danielezhu]

This is an incompatible change we should keep tabs on. It's less "dangerous" since we're augmenting the list instead of deleting elements from it, but at the end of the day, we're changing the value of a constant.

[oyangz]

(non-blocking): should we move the BertScore class out of summarization accuracy metrics since it’s used for multiple eval algos now?

[danielezhu]

(non-blocking): should we move the BertScore class out of summarization accuracy metrics since it’s used for multiple eval algos now?

If we do, it will be a breaking change that we should keep track of prior to the next release.