This seems pretty efficient (approx 12M write/s as opposed to 2M with
the DefaultCounterService). N.B. there is no need to change most of
the rest of the metrics stuff because metrics are write-often, read-
seldom, so we don't need high performance reads as much.

The Spring Integration configuration and Dropwizard support has changed
a bit. Functionally very similar and probably opaque to users, but now
the messaging operates as an Exporter on a @Scheduled method, and
Dropwizard is a replacement [Gauge,Counter]Service.

Metrics are all
collected live in-memory (and can be very fast with Java 8), buffered
there and shipped out to a MessageChannel (if one exists with id
"metricsChannel") in a background thread.

We can still use Java 8 library APIs (like LongAdder) but to compile
to java 7 compatible byte code we have to forgo the use of lambdas :-(
and shorthand generics (<>).

Fixes gh-2682, fixes gh-2513 (for Java 8 and Dropwizard users).