We use "monotonic" time-stamps for the dead_since field;
that breaks on system reboots, in which case we reset the stats.
(if the server was categorized as dead)

If the server times out afterwards, we'd fail the condition
`cur_state.consecutive_timeouts == old_state.consecutive_timeouts`
so its stats would not update.  Therefore we'd get stuck forever
in a state where the unusable server has high priority (no_rtt_info).

This commit changes a bit more than was necessary to fix this,
including precision of the stats (in some cases).