Fix bug where state_group tables got corrupted

This is due to the fact that we prefilled caches using txn.call_after,
which always gets called including on error.

We fix this by making txn.call_after only fire when a transaction
completes successfully, which is what we want most of the time anyway.