CandidateGenerator.getRandomIterationOrder is too slow on large cluster

In the actual implementation classes of CandidateGenerator, now we just random select a start point and then iterate sequentially, instead of using the old way, where we will create a big array to hold all the integers in [0, num_regions_in_cluster), shuffle the array, and then iterate on the array. The new implementation is 'random' enough as every time we just select one candidate. The problem for the old implementation is that, it will create an array every time when we want to get a candidate, if we have tens of thousands regions, we will create an array with tens of thousands length everytime, which causes big GC pressure and slow down the balancer execution.