It's not possible to start a worker from outside the dispatcher
without the receiver's client socket.
Since client_info is already passed around,
it's the logical place to add the socket to.
This way we're gonna be able to start a worker inside the allocate
request handler code.

client_info.port is port on which we receive datagrams from the client.
Therefore, we can't use this client_info instance to create
a reserved worker for the new allocation,
as the client will connect from different port.
We have to rely on dispatcher to create the new worker process,
but at the same time store the reservation somewhere,
so that it can be picked up by that new worker when it reads
the RESERVATION-TOKEN attribute coming with an allocate request.

This dramatically reduces the overhead (approx. 3x less space consumed per reservation).

Until now the log was tied to the server's listening port.
This does not work well with the test helpers we have in place,
and is not conformant with the RFC:

> [the client] can use a different client IP address and
> port, a different transport protocol, and even different server IP
> address and port [when requesting an allocation with a RESERVATION-TOKEN]