It reuses a standard mechanism, is less verbose and avoids including
code that is never executed.

Queue.get() accepts any non-negative number, so 0 works. It's also not
arbitrary like 1 was, and the test suite now takes one less second to
run!

It was introduced in urllib3/urllib3#875, but
since urllib3/urllib3#1495 twine is no longer a
dev requirement.

Instead, install it on demand like twine.

The idea is to make sure that we'll get the pytest version we want: the
last that supports Python 2. Using a different version in Python 2 and
Python 3 is possible, but is likely to lead to trouble in the future.

It turns out that without turning this one it's difficult to list
exactly the tests that failed.

The default pytest tracebacks are really verbose: for each function in
the stack they include the whole function until the exception, and the
representation of its parameters.

On the other hand, native tracebacks are way more concise, showing only
two lines for each function and exception chains.

The benefits of switching to native tracebacks:

 * The useful information fits in one screen
 * Python trains us to read those tracebacks from day one
 * We avoid the verbosity of showing the source code: we can already
   see it in GitHub and our editors
 * While we lose the information about parameters, I've never found this
   information to be useful

I've been using this configuration for a few months now and never looked
back!

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30

…om-master-2019-10-30