A batched save fires MediaRequestSubscriber.afterUpdate for every request
concurrently, and each opens a nested transaction on the same query runner.
On Postgres the concurrent savepoints collide, aborting the outer transaction
and returning the connection to the pool in a broken state.

fix #3365