MathML corpus statistics

[dginev]

Requested by the MathML 4 effort, I added a small example script that extracts a element-attribute-value footprint over the <math> elements of a given llamapun corpus.

Running this over arXiv, will leave the PR open until I vet that the final data looks good.

Request issue with details here:
w3c/mathml#55

Aside: I also simplified the corpus_node_model example -- which still doesn't have any practical purpose -- to use a Result-returning main()

[dginev]

Happily, the DLMF can be seen as a llamapun corpus if we just filter over the .html5 files, and ignore all other auxiliary content. Then the MathML statistics script can also be run over the DLMF, which takes the pleasing 15 seconds. So far so good.

[dginev]

As a side note, this PR also motivated looking into multi-threading the rust libxml wrapper, (see KWARC/rust-libxml#47 ), to speed up this type of traversal on modern CPUs. Having the multi-threaded libxml may also boost us in creating DNM at scale faster, but that's something for another day.

Right now traversing arxiv takes 70-75 hours (for 1.2 million .html files) on my spinning HDD storage, and the bottleneck is not at IO, while also using only a single CPU. So definitely room to squeeze more performance out.

[dginev]

I have now generated the arXiv report, and generally feel this is a good state for the stats example, so merging here.