- Ceil/floor compatibility with pre OTP 20 versions

- No named table even for concurrent access

  The table id is stored in the #hist{} struct (which is shared with all
  updating processes) anyway, so no need for a registered name.  This is
  a minor optimisation to avoid name->tid lookup. (Working with tids was
  slightly optimised in OTP 20 when they became references instead of
  integers)

- One table and one entry per histogram

  Create one table and one entry when initialising the histogram.
  Delete the whole table when deleting the histogram.

Compatibility with hdr_histogram_erl NIF

- Use erlang:make_tuple when reseting row

  This avoids creating a huge temporary list. Reset is called relatively
  infrequently, so this change does not have too big impact.

- OPT: calculate bucket/subbucket index from index

- Collect multiple stats in single iteration

Store pid/regname of event handler process in application env.

A message in the form `{snapshost, NowTime}` is sent to the event
handler when a snaphost is created.

The testcases subscribe themselves as event handlers making them
synchornous and bit less prone to timing issues.

So far this is only used for tests but could be the start of a
websocket-based notification system.