* Use PyPDF2 (new dependency) to loop over images attached to a PDF
  this allows to detect and raise an error whenever there's >1
  image/page (we cannot treat this correctly anyway)
* rewrite helper functions to deal with image objects rather than reading
  from disk
* This also means we don't need pdfimages anymore

This makes 'process_pdf' easier to use from a Python shell, which
helps with debugging

Previously we were using zero-based indexing, which is confusing.