Add a sample of two ancestor with each node

This allows readers to efficiently get ancestors of nodes with low indegree
(ie. most revisions), as it avoids a random access / API call.