Make PyMuPDF always log to stderr

[apyrgio]

Problem

PyMuPDF has made the design decision to log all messages by default to stdout (see pymupdf/PyMuPDF#3135). This is a problem for us for two reasons:

First, in the document to pixels phase, we assume that the stdout of the conversion process contains only the pixels of the rasterized document (plus some page/size info). If a library that we use also logs to stdout, then this will corrupt our pixel stream, and our conversion will fail.

This is actually biting us in practice. We have a malformed file (no-pages-types.pdf) in our large test set that causes PyMuPDF to log this error to stdout:

MuPDF error: format error: non-page object in page tree

Tip

You can reproduce it with:

$ python3 -c 'import fitz; list(fitz.Document("no-pages-types.pdf").pages())' 2>/dev/null   

The error above is mixed with the pixel stream, and Dangerzone fails.

Second, in the pixels to PDF phase, we had a case (#700) where PyMuPDF was logging to stdout, from which we expect to read conversion progress reports in JSON format. We circumvented this using Python's contextlib.redirect_stdout(), but we lose error messages this way.

Solution

A solution to the above problem that works for all configurations is to take advantage of the environment variables that PyMuPDF devs have added to control logging. An important detail here is that we have to use these environment variables in our dangerzone.conversion.* modules, before we load the fitz module, which is ugly, but doable.