Some buggy documents can produce a huge stderr output because of parsing
errors or whatever. We could give a file directly to Popen to write the
stderr file, but we only want to create the file when there's output,
because it's what we use to know whether the command produced output or
not. So, instead of buffering the whole output and then write it to the
file, now we read from the pipe while the command is running, writing
the output in chunks to the file. This improves a lot the memory
consumption when running some tests.