Change the logic of who calls callbacks:

  eb_device_close kills outstanding callbacks, not eb_socket_close.
... to do this, response tracked cycles remember their device.
... and on device close it filters the entire list of awaiting responses.

The reasoning behind this change:
  the user probably has handles to the device in the callback
  if he closes the device, the callbacks might run later with a bad pointer
  instead => kill those callbacks as part of closing the device.