Use the system's value for HZ (100 by default), which is the unit that
cgroups cpuacct statistics are using.

Actually register the right prometheus Handler. Also make the server
wait for a metrics collection cycle to complete before starting the
HTTP server, so we never serve missing metrics.