Added score tie-breaking to BM25PrfReranker

[emmileaf]

In this PR:

• Added tie-breaking to second retrieval in BM25PrfReranker
• Fixes MS MARCO passage regression errors: BM25prf gives non-deterministic results #774

@lintool before merging this, could you help with sanity checking on a different machine if it's not too much trouble? I re-ran indexing + retrieval twice on tuna to check and there were no changes to the regression numbers, but non-determinism issues might be easier to slip by...

[matthew-z]

I also added that snippet in my local, and the results are now deterministic

[lintool]

@emmileaf can you update the regression numbers and I'll also see if they match? Thanks!

[emmileaf]

The regression numbers didn't change when I re-ran after the fix, so hopefully it should be able to pass with the current ones?

[lintool]

On tuna, I'm still getting:

2019-08-12 19:27:45,991 - regression_test - INFO - {"actual": 0.1582, "collection": "msmarco-passage", "expected": 0.1582, "metric": "map", "model": "bm25-tuned+prf", "topic": "[MS MARCO Passage Ranking: Dev Queries](https://github.com/microsoft/MSMARCO-Passage-Ranking)"}

and

2019-08-12 21:51:51,442 - regression_test - ERROR - !!!!!{"actual": 0.1583, "collection": "msmarco-passage", "expected": 0.1582, "metric": "map", "model": "bm25-tuned+prf", "topic": "[MS MARCO Passage Ranking: Dev Queries](https://github.com/microsoft/MSMARCO-Passage-Ranking)"}!!!!!

Across two different trials. Untuned BM25 seems to be okay, though. Any ideas?

[emmileaf]

I can't seem to reproduce this error on tuna anymore…perhaps try with #776 instead? We both pushed a fix at the same time, but it looks like @matthew-z made a few additional small changes in that PR?

[lintool]

I reran a two times on two different machines and am now getting consistent results... I must have been inadvertently running old code? Merging.