txn_complete used to handle all the transaction outcomes:
- manual rollback;
- error at WAL write;
- successful WAL write and commit;
- successful WAL write and wait for synchronization with replicas.

The code became a mess after synchronous replication was
introduced. This patch splits txn_complete's code into multiple
pieces.

Now WAL write success and fail are handled by
txn_on_journal_write() exclusively. It also runs the WAL write
triggers. It was very strange to call them from txn_complete().

txn_on_journal_write() also checks if the transaction is
synchronous, and if it is not, it completes it with
txn_complete_success() whose code is simple now, and only works
on committing the changes.

In case of fail the transaction always ends up in
txn_complete_fail().

These success and fail functions are now used by the limbo as
well. It appeared all the places finishing a transaction always
know if they want to fail it or complete successfully.

This should also remove a few ifs from the hot code of transaction
commit.

The patch simplifies the code in order to fix the false warning
about too long WAL write for synchronous transactions, which is
printed not at WAL write now, but at commit. These two events are
far from each other for synchro requests.

Part of #5139