[Metrics] Fairseq BLEU Re-Implementation

[klshuster]

Patch description
Prior experiments for e.g. dodecadialogue have utilized fairseq's tokenized bleu scores for measuring BLEU. From inspection of the fairseq generation script, it becomes apparent that they calculate BLEU on a macro level, considering all predictions and references together. Our original implementation did not take into account the brevity score, which is a total across all predictions/references, and thus our scores were misaligned.

I've thus re-implemented fairseq's bleu scoring with ParlAI's metrics system. I have verified that the computed BLEU score for e.g. the command referenced in #3473 yields nearly the same BLEU score as reported in the paper.

Testing steps
Added tests to CI:

$ pytest -k TestFairseqBleuMetric -v
====test session starts ====
...
collected 477 items / 476 deselected / 1 selected

test_metrics.py::TestFairseqBleuMetric::test_scorer PASSED                                                                                                                     [100%]

====slowest 10 durations ====
0.12s call     tests/test_metrics.py::TestFairseqBleuMetric::test_scorer

(2 durations < 0.005s hidden.  Use -vv to show these durations.)
====1 passed, 476 deselected, 3 warnings in 5.13s ====

Additionally:

$ parlai eval_model -mf zoo:dodecadialogue/empathetic_dialogues_ft/model -t empathetic_dialogues --skip-generation false -dt test --metrics ppl,bleu --inference beam --beam-size 10 --beam-min-length 5 --beam-block-ngram 3 --beam-context-block-ngram -1 --compute-tokenized-bleu true -bs 32
.
.
.
 fairseq_bleu1  fairseq_bleu2  fairseq_bleu3  fairseq_bleu4 
  15.61         8.018           5.159            3.485

[stephenroller]

Great sleuthing