Rework how we handle multi-occurrence images

Multiple occurrences, possibly with different transformations, are now
listed as separate images, the expensive image decoding is only done
once though by using a document-wide image data map. This fixes a few
correctness issues (e.g. when listing images for a page sub-rect), and
more importantly brings this a big step closer to the discussed Qt image
access API in Poppler.

This will possibly duplicate barcode decoding effort though, but that's
something to be handled one layer up in the extractor, by using a object
id based mapping like we use for the image decoding.