The code for generating unique_together indexes failed to set a valid

name, which was fine so long as that index wasn't being set twice in the

same transaction.
With this bug, setting a unique_together and then changing it would

result in the SQL for index creation happening more than once.
It also didn't filter out no-op index creation/removals, which was

actually a problem for Postgres.
Now, we do a better job of filtering out unnecessary index operations

and also dealing with indexes that were created within that transaction.

Unit tests pass.
I hit this problem in production. This patch fixed it.