[Store] Shrink metadata maps after eviction cycles (#3576)

unordered_map::erase never reduces the bucket array, so a shard that
once held millions of keys keeps its high-water bucket memory forever
(master_service.cpp has no rehash/shrink call at all). After a batch
eviction, rehash maps whose live size dropped below a quarter of their
bucket count. Part of the RSS findings in #3452.