Catch up with change to Cluster configuration.

A hash is more ops-friendly (notably it accommodates storing the whole config natively in consul's kv store using keys like "InstanceTypes/m1.large/price") and aligns with the existing Nodemanager config format. It also makes it more obvious that the instance type name has to be unique.

During a transition period programs that read the cluster config file should accommodate both array and hash forms, so ops can migrate their config files without stopping the world.