We aren't changing `label` or `verbose_name` so let's use class inheritance instead

ref https://github.com/django-oscar/django-oscar/issues/3080

- Move Oscar app configs into Django app config classes
- Rename Oscar/Django apps' app config module
- Remove "oscar.get_core_apps()" and its app overriding
- Remove "oscar.OSCAR_CORE_APPS"
- Port app forking code to work with Django app registry
- Port class loading code to use app configs to get app properties

If a phone_number field is required, and empty, then PhoneNumberMixin.clean() fails with a key error because the field does not exist in cleaned_data.

This patch fixes the clean() method to only attempt to clean a field if it exists in cleaned_data. If it doesn't then allow Django to raise a validation error for a missing required field.

Also added more comprehensive tests for the validation logic in the mixin.

Previously, validation did not pass when current locale is "en-gb" and URL is "/de/catalogue/" and visa versa due to default Django URL resolvers behaviour  - they do not take into account all locales, only current.

phone_number_fields is more correct grammatically.

Add to the dosctring for the class, and display it in the documentation.

With this change phone numbers (any qty) can be added as `ModelForm`
property `phone_numbers_fields` - dict where "keys" are names of
phone numbers fields and "values" are attributes of fields.

* `PhoneNumberMixin` clightly reworked and moved
from `oscar.views.generic` to `oscar.forms.mixins`
* Couple tests added (based on previous tests)

Fixes #2414.

Django 1.9 introduced the AUTH_PASSWORD_VALIDATORS setting which allows for
more flexible password validation. Switch to using that when it is available,
with a fallback to Oscar's previous behaviour in Django 1.8 or when this
setting is empty.

This was removed from the parent class way back in Django 1.4, and was
not being used at all here. It existed only for backwards compatibility.

It seems that migrations are unicode in Django 1.11

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.

* Tidy implementation of validator
* Tests improved and moved to tests/integration (as they call resolve
  internally)