Add benchmarking for table extraction

[jamie-lemon]

We want to see how well the solutions do when extracting tables from PDFs, let's source some files and add this.

[hashiromer]

Okay so I believe finding or curating dataset should not be difficult but I will need to think about a good evaluation strategy for table extraction. In any case, i will update you in a few days.

[hashiromer]

@jamie-lemon

Which table format does pymupdf-layout support? Simple tables or full HTML tables?

I am asking because the evaluation criteria differs for both. HTML tables are strictly much more expressive compared to simple tables and are widely used. A lot of tables cannot be even be represented correctly in a simple markdown table.

FYI, markdown itself doesn't have a standard, there are many flavors of markdowns. Some fully support rendering HTML tables and some don't. For LLM usage, the flavor doesn't matter at all because LLMs don't render markdowns, they read them so its better to produce the most expressive markdown flavor which supports full HTML tables.

[bonfab]

Complex tables with nesting and merged cells are not uncommon. Supporting HTML tables would therefore be very helpful.