We had two problems that were potentially causing the existence of

duplicate LocalSiteProfile models. The first of these is that this

wasn't using ConcurrencyManager, which handles integrity issues across

separate threads (or servers). The second, more difficult one, is that

some databases (particularly MySQL) were ignoring our unique_together

constraint when one of the relations was NULL. This meant it was

possible to end up with multiple LocalSiteProfile objects with

local_site=None.
This change moves us over to ConcurrencyManager, and implements a

check for the duplicate case when attempting to use

User.get_site_profile. If we do find duplicates, they'll be

consolidated into a single one before returning.

Ran unit tests.