Make block distance dynamic in multi-columns articles

[lfoppiano]

There is a piece of code:

https://github.com/kermitt2/grobid/blob/92ea31edc391c56f6fd1eab61a16aaab5fda960c/grobid-core/src/main/java/org/grobid/core/document/Document.java#L1276

that decided whether blocks around a figure caption (something starting with fig, and tagged as <figure> is "close" enough to it (15px fixed distance) so that consider it or not as part of the figure. Before #1266, certain part of the document were discarded, however now are added as paragraphs.

Now, this works in many cases, but it relies on the fact that the blocks distance is consistent.

in this article, for example, the figure caption blocks are correctly recognized:

Figure 1 is correctly assembled:

Figure 2 is not, because the blocks dimensions are too "far" (15.16px > 15px):

By looking at few examples, if and ends near the column, it will make the block larger, and therefore, close to the other...

I did relax the rule of the 15px when the two blocks are at the same height, but is not a solution that will cover all cases.

any comment is welcome 😄

[lfoppiano]

Here a better solution: #683 (comment)

[coveralls]

coverage: 40.657% (+0.2%) from 40.422%
when pulling af75090 on bugfix/adjust-block-distance-when-multi-columns
into 49df55c on master.