kubeletplugin: join the socket cleanup error, not the primary one twice

Both cleanup paths in endpoint.go passed the primary error to errors.Join
twice instead of joining removeErr, so a failed socket removal was thrown
away.

In listen that is visible to a driver: the error travels through
startGRPCServer to kubeletplugin.Start, and when the stale-socket removal
also fails it reads as the listen error printed twice, without the removal
failure that explains why the bind could not happen.

In unixListener.Close it is latent. The listener normally closes cleanly,
which made the call errors.Join(nil, nil), and that is nil, so Close
reported success with the socket still on disk. Nothing observes it today,
because the listener is handed to grpc.Server.Serve, which drops the error
from closing it, and unixListener is unexported. It is fixed because the
code does not do what it says.

The test covers Close only: listen removes the socket before binding and
returns early if that fails, so its second removal cannot be made to fail
without changing the code.