It should be a lot faster. Because it's a new required tool,
we must shuffle couple things around.

While using debbindiff with file system images mounted with different
loop devices, stat shows that those loop devices are indeed
different. This clutters the report if you are comparing many files and
unable to see important points.

Usually those devices does not matter when comparing files for
reproducible build, so disable the device diff.

But to get them all we would need to parse `file` output in text mode.
Let's see later.

Closes: #782551

Closes: #782253

Zip member filenames may contain "../", which
is ignored while extracting. ZipFile.extract returns
the actual path where the file is extracted to.

GNU diff is unable to cope with very large files as its memory usage
grows too much. So we now have a maximum input size to avoid getting OOM.

Convert msgunfmt as an example.

This paves the way for huge memory savings.
Convert xxd as an example.

Difference() now takes an unified diff directly. Computing the diff is
moved to a new static method from_content() which returns None when there are
no differences.

This paves the way for passing file descriptors to from_content() to avoid
loading entire outputs in memory.

This reverts commit 303a31e9.