Using magic and doing smart comparisons is slower than comparing bytes
directly. Let's assume most files will actually be identical and start
comparing small files directly. If they match, we can skip the smart
bits entirely.

This lead to drastic run time improvements when comparing source tarballs
with many small files (e.g. linux-source).

Closes: #773573

Even when we fallback to binary comparison.

We actually want to use the path given by the caller. We now properly
display the path of a file when it is in an archive.

D'oh!

No changes required.

Closes: #764470

Seems I did not properly read the warning in subprocess documentation:
“Do not use stdout=PIPE or stderr=PIPE with this function as that can deadlock
based on the child process output volume”.

And indeed, we had a deadlock on files with very long lines. So let's switch
to use Popen and communicate().

Closes: #764131, #764420

Some jar files contained no apparant differences in plain `zipinfo`,
but showed some in `zipinfo -v`. The former is more readable, so still try
to get some result with it first.

Thanks to Hans-Christoph Steiner.

Thanks to Hans-Christoph Steiner for the suggestion.

At the request of Holger, we just don't create the HTML file if there is no
differences to report.

difflib is unbearably slow for any non-trivial output. Calling out to vim is
much faster and the output is still quite nice. Crazy trick, though.