NoValidConnectionsError not raised when socket.getaddrinfo() fails as a direct cause of client.connect()

Documentation for the NoValidConnectionsError states that:

This exception class wraps multiple “real” underlying connection errors, all of which represent failed connection attempts. Because these errors are not guaranteed to all be of the same error type (i.e. different errno, socket.error subclass, message, etc) we expose a single unified error message and a None errno so that instances of this class match most normal handling of socket.error objects.

IOW it wraps all various kinds of socket errors. However, this doesn't include calls to getaddrinfo in the _self._families_and_addresses method. IMO issuing getaddr is an inherent part of establishing connection and hence socket.gaierror exceptions should be wrapped like any other exception coming from socket.socket. This forces library users to handle both types of errors not just NoValidConnectionsError which is IMO in contradiction to the docs linked above since from client's POV both exceptions came from the same library (paramiko in this case).

To provide a bit more background, I'm using paramiko to test whether a fresh VM I created is ready to be connected to over SSH. It takes a little while until the VM reaches a state where the network is up and the VM got a DHCP lease, so naturally getaddr must fail, but that is expected, it doesn't change anything on the fact, that gaierror caused as part of client.connect should not be wrapped with NoValidConnectionsError.

Start at Paramiko's client.connect flow and the _self._families_and_addresses method named in the report. Compare the documented NoValidConnectionsError behavior with socket.getaddrinfo() failures, then add regression coverage showing the intended exception from client.connect and run the relevant test suite.