thread OperationContext* down through ClusterClientCursor's next() and kill() methods

This removes the need for the AsyncResultsMerger to store an OperationContext*, which it uses to schedule finds, getMore's, and killCursors on remote shards.

This makes it easier to deal with the RouterStageMerge being split into two parts (sending the finds to establish cursors, and sending the getMore's to pull and merge results), where both parts need to access the OperationContext*.