Adds the option to run on a whole document at a time

[dirkgr]

Other changes:

• Annotations are now in the fields "bff_duplicate_spans" and "bff_contained_ngram_count". Those names have changed.
• bff will always create the "bff_contained_ngram_count" field, even when running deduplication.

[dirkgr]

To be sure, I don't know what the point of this is. It would be insane to run deduplication this way. It would be borderline insane to even count ngrams this way (i.e., with --annotate-only). And it's not implementing the full GPT2 spec from #3 either, which is already insane all by itself.

So I'm just writing the code and I trust that you know what you're doing with it.

[IanMagnusson]

This looks good to me! For posterity, we sorted out the two issues: 1) fixing the update bloom filter flag so it actually works, 2) using filter threshold of 1.0 when priming the bloom filter to make sure all ngrams are included without any skipping.