We have some custom SQL that has to run to compute the local-site specific ID

for review requests. Previously, we called transaction.commit() after doing

this, which works fine in production, but is problematic when running under

Django 1.8's unit test class, which puts the entire test case into an atomic

block.
Instead of manually calling transaction.commit(), I've wrapped the SQL in its

own atomic block. In production, this acts the same as calling commit()

afterwards, but in the test case (where it's a nested atomic block), it runs as

a "save point", doing the right thing instead of crashing.

Ran unit tests on 1.6 and 1.8.