Do not store socket_name in a reference. Instead, provide two calls:
[connect_external] that takes the socket name as an argument and
[connect_internal] that starts a dedicated server on a fresh socket.
If [Prove_client.send_request] is called when no connection is
established, it calls [connect_internal].

Rationale: the only reason to have a reference for socket_name is
to delay the connection request (for example, until the first call
to [Prove_client.send_request]). However, if the server is already
running when we set [socket_name], it does not cost us anything to
establish this connection right away. This is what [connect_external]
is for. Otherwise, if the server is not running yet, the API client
must ensure that it will be started before the first [send_request],
or else the connection request fails. This requires synchronisation
between the entity that starts the server and the Why3 process. In
this case, the API client should just call [connect_external] once
it knows that the server is up and running.

TODO:
- what is the semantics of [disconnect]? How should we handle the
  tasks in progress?
- what is the semantics of [max_prover_running] when working with
  an external server?
- how should we handle server-side disconnects (if at all)?