remove spike output callback from communicator

* make a new gathered_vector type for storing a "distributed" vector
  that has been gathered onto a single node
    * stores the vector
    * stores the partition information
* split communicator::exchange() into two phases
    * exchange() which simply exchanges spikes
    * make_event_queues() which generates local queues from spikes
* now callbacks for both local and global spike store can be
  performed in the model
* and extra meta data about the partition is available for more
  advanced parallel global storage schemes.