One PairFact repo for ACL code review

[lihebi]

This is the up-to-date code & results for PairFact, including factcc,qags,frank datasets. It is all-in-one repo ready for ACL code review. Changes in this PR:

1. add factcc,qags,frank datasets (from PR (WIP) add factcc, frank, qags datasets #6)
2. use pooled-level instead of summary-level for factcc & qags, because there's only one system for each docID, thus cannot compute correlation.
3. run on both cnndm & xsum
4. move bertscore & mnli metrics into DocAsRef_0
5. add generated results to this repo
6. skip experiments if the corresponding result files exist.
7. add some description to README.

TODO: anonymize it.

The important file is factcc.py and eval.py (the entry point). Other files are mostly copied from DocAsRef_0 repo.

Current results:

	frank-cnndm	frank-xsum	qags-cnndm	qags-xsum	factcc
system-level	done	done	done	done	done
summary-level	done	done	-	-	-
pooled-level	?	?	done	done	done

Note: the pooled-level collect all results for all docIds, and compute correlations for the entire vectors. The reason is qags and factcc has only one system per docID, thus cannot do summary-level (stats.scipy.correlation throws error on vectors of size 1).

Additional missing results (currently running):

• qags-xsum-pooled Summary-Level
• frank-xsum Summary-Level

[lihebi]

Update: add classification results in the results-classification folder. Include the following thresholds: 0.3, 0.4, 0.5, 0.6, 0.7