Transaction support

We need to support transactions. This is needed mainly to make stored procedures not only atomic, but also durable and consistent.
Since our protocol is fully asynchronous, we need to make sure that transactions do not hold state
on the server. The transaction should be present in the WAL (and, possibly, WAL write buffers), but the rest needs to belong to the client.
What's a transaction identifier? We need to come up with one. It needs to be part of the request packet.

We don't need transaction identifiers as long as we do not expose transactions in the protocol -- aka we only allow transactions in stored procedures.
And we don't need interactive transactions much nowadays.