That's probably more or less the same thing in terms of what actually happens in PostgreSQL.
All things considered, an actual database probably gives you the least gray hair, but with some careful test setup I have had good success using the savepoint/rollback trick (and it trivially supports nested fixtures as well).
With multiple databases you get the advantage of being able to run tests in parallel, so even though they might be the same under the hood they offer interfaces that suit quite different use cases
