monitor status collection should be done in parallel

The master collects statistics every few seconds from all the tservers so that it has up-to-date statistics to hand to the balancer. It does this one server at a time. On a very large cluster, the status collection can take a long time, causing the monitor to show "blocky" statistics in the graphs, and long "last contact" times in the tablet server lists.

A small thread pool could be used to speed up this simple work.