While you may or may not be able to solve this particular problem, I'd suggest that avoiding it entirely might be a better option. As you've seen, depending on a transaction to guarantee that your database is in a particular state doesn't always work. Also, because you are using multiple connections to the DB, you've automatically promoted any transactions that do occur to distributed transactions -- a subtle distinction, perhaps, but it changes the nature of the test. You may end up writing code to overcome the particular limitations of distributed transactions that wouldn't otherwise have been needed.
A better strategy would be -- for unit tests, anyway -- to mock out the database dependency, using in-memory mock or fake objects in place of the database. I've done something similar for LINQ to SQL (see my blog entry on the subject) For integration tests, I think you are better off using a test instance and writing set up code that reinitializes the state of the DB to known values before each test than introducing an extra transaction to clean things up. That way if your clean up code fails in a test, it won't affect other tests being run.