Optimize ReservoirSegmentSampler.getRandomBalancerSegmentHolder()

The current algorithm is `O(N_SEGMENTS_IN_CLUSTER)`, it can be `O(N_SERVERS_IN_CLUSTER) + O(log(N_MAX_SEGMENTS_ON_A_SERVER)` via reservoir choice of a server first using populations of servers, then a random segment can be chosen on a server using [this hack](https://stackoverflow.com/questions/12385284/how-to-select-a-random-key-from-a-hashmap-in-java/54893402#54893402).

Locate ReservoirSegmentSampler.getRandomBalancerSegmentHolder() and inspect its current segment-selection path. Compare the existing complexity with the issue's proposed server-first reservoir selection and per-server random segment lookup. Done means the method achieves the stated complexity improvement while preserving random selection behavior; no test file is named in the issue.

Written by the indexing model from the issue text.