Introduce support for PDFs

[yabramuvdi]

Feature request

The idea (discussed in the Discord server with @lhoestq ) is to have a Pdf type like Image/Audio/Video. For example Video was recently added and contains how to decode a video file encoded in a dictionary like {"path": ..., "bytes": ...} as a VideoReader using decord. We want to do the same with pdf and get a pypdfium2.PdfDocument.

Motivation

In many cases PDFs contain very valuable information beyond text (e.g. images, figures). Support for PDFs would help create datasets where all the information is preserved.

Your contribution

I can start the implementation of the Pdf type :)

[yabramuvdi]

#self-assign

[lhoestq]

Awesome ! Let me know if you have any question or if I can help :)

cc @AndreaFrancis as well for viz

[lhoestq]

Other candidates libraries for the Pdf type: PyMuPDF pypdf and pdfplumber

EDIT: Pymupdf looks like a good choice when it comes to maturity + performance + versatility BUT the license is maybe an issue, and pypdf, pypdfium2 or pdfplumber are good options imo

[AndreaFrancis]

Related to #7058

[yabramuvdi]

PyMuPDF is AGPL licensed, so we can't use it. I will move forward with pdfplumber.

[yabramuvdi]

Hi both! I have made a pull request with a first basic implementation of the Pdf feature. I followed closely what I saw on the Video and Image features. It is my first time contributing so any comments are very welcomed. I think it would be useful to outline together what additional things we can implement (e.g. enabling parsing of the pdf). Thanks :)