The primary change is to fix a memory leak reported in
Netflix/spectator#264. Each time a gauge was updated it
was creating a new registration and because the map holds
a strong reference these would never get collected.
Further, the aggregate value created by the multiple
registrations was not correct.

In addition I added some test cases around the various
inputs and checked that the results were reflected as
expected in the registry. I noticed the timer values
had a unit of milliseconds, but it isn't immediately
clear if the reported value can ever less than 1.0. The
conversion to long is now delayed until after converting
to nanoseconds so duration values less than 1.0 will now
work instead of just recording 0.

For the histogram I changed to just using a cast to `long`
to avoid boxing to a `Double`. As an FYI for the future,
there is a DoubleDistributionSummary we have experimented
with in spectator-ext-sandbox that might be more appropriate
for this use-case.