When rb-site upgrade determines it needs to update settings, it first

writes a new settings_local.py file and then reloads Django's internal

settings.
The way it reloaded settings wasn't very safe, and that's been fixed up,

but the entire block of logic for reloading settings is problematic in

unit tests. While the unit tests for settings migration work fine, they

impact other unit tests. This happens by removing the cached

settings_local module in Python and then immediately triggering a

reload.
Since the unit tests are working with test versions of this file, what

we end up doing is finishing the test suite with django.conf.settings

wrapping a test version of settings_local.py. Future unit tests are

then sensitive to this, and blow up.
The code has been updated to only perform a settings reload if we know

that the settings_local in cache is the same one containing settings

we're migrating. This will be true when running rb-site upgrade, but

won't be when performing unit tests.

All unit tests that failed due to this issue now pass.