A scrape counts the resources where they already are

[bjmeetsfo]

Every Prometheus scrape of the metaserver listed each server, each proxy,
each namespace, each table and each proxy group -- five calls, five times
taking the lock, five copies of a collection -- and then used the lengths
and the states and threw the copies away.

Listing a server copies it whole, and a server carries its shard-load list
and its shard-state list: three strings for every shard it holds. A cluster
of thirty-two datanodes holding a thousand shards each had ninety-six
thousand strings copied to answer how many servers are frozen, on every
scrape, on an interval, forever.

The totals were already gathered where the state is held -- MetaStats
carries them. It now carries the per-state breakdown too, counted in the
same pass, and the per-server and per-proxy numbers the scrape reports come
from a row that has those numbers and nothing else.

servers x shards each     before      after
          8 x    250      372 us    19.2 us
         16 x    500    2 360 us    29.9 us
         32 x  1 000    8 559 us    44.6 us

A hundred and ninety times faster at the largest, measured before and after
back to back. It also stops growing with the shards: sixteen times the
shard states cost 2.3 times as much rather than 23 times.

A test registers a server in each state and checks the counts read from the
stats are the counts listing them would have produced.