Update of the header model

[kermitt2]

The header model requires a big refresh to be consistent with all the evolutions of GROBID during the three last years.

• modify pdf2xml for having blocks in reading order at least for the header segment, see fork https://github.com/kermitt2/pdf2xml
• use only the Segmentation model for getting the header segment for both training data and files to be processed
• modify the training format to be consistent with the rest of GROBID's training
• add some new features (line indent, centered, ...)
• fix some issues with the training files: see Duplicate ID in Header Corpus #135, + 2 PDF files in training set has no content.
  Use full file name as identifier and check for duplicates
• review carefully modified training data format
• regression test and hopefully some improvements ;)

[kermitt2]

This is work in progress with branch modified-pdf2xml...

[lfoppiano]

I'm adding a PDF example as test case for this extension. While the text is correctly extracted from pdfalto, using -readingOrder option, the model is mixing up (I suppose) the various streams.

I guess after updating the header model this pdf should be extracted normally.

https://arxiv.org/ftp/cond-mat/papers/0111/0111388.pdf