The idea of splitting integration from unittests is good in theory
but leads to a lot of mental overhead. Besides whenever a tests
interacts with a database it isn't a unittest anyway.

Better to use filenames which are commonly used in the python word then
custom ones.

In Django 1.9 the decorator from six is used so simplify the test to
just check for a __unicode__ method

The changes should be obvious and non-controversial. Under Django 1.7+,
the AppConfig is used to register signal receivers instead of just
registering then when models are registered.

Working on sets causes a random order of returned fields while using
existing_user_fields as reported in django-oscar/django-oscar#1597.
Changed to work only on lists, which remain correct
order.

Closes #1605.
Closes #1597.

We only needed it because Django 1.4 shipped with a pretty old version
of six. Support for that has been removed, and Django 1.5 ships with six
1.6.1, which is more current than we required.

This nicely avoids an issue with django-extra-views pinning a six
version which caused the sandbox build to fail:
https://travis-ci.org/tangentlabs/django-oscar/jobs/32223978#L971

This commit adds the python_2_unicode_compatible decorator to all models
that define a text representation. Furthermore, it adds a test to ensure
that we never forget to do that.

It is explained in details in the Django docs:
https://docs.djangoproject.com/en/dev/topics/python3/#str-and-unicode-methods

This required extending the UnicodeCSVWriter to be able to deal with
file-like objects as well. This is used to create downloadable CSV files
in memory.