Don't start rolling update with seed nodes

[a1exsh]

Currently the create_cluster code selects seed nodes from the lowest private IP addresses, and update_cluster sorts the instances to update by private IP.

The overall effect is that when performing a rolling update, some seed nodes are updated first, which can result in additional trouble if the update doesn't work as expected: the seed nodes have different startup behavior and replacing them in case of major problems is harder than with normal nodes.