fix(threadstats): ensure MetricsAggregator is threadsafe

[TheKevJames]

As described in #142, the MetricsAggregator is not threadsafe under Python 3. This is a regression caused by a change in how the .keys() method works under Python 2 vs Python 3: in Python 3, calling .keys() returns a view rather than a copy of a list.

The solution here is to use a reentry lock around the modifiers and accesses to self._metrics, in order to ensure that any concurrent calls to MetricsAggregator.flush() and MetricsAggregator.add_point() have no race conditions.

Note that the stack trace which this changeset fixes is:

RuntimeError: dictionary changed size during iteration
  File "datadog/threadstats/base.py", line 307, in flush
    metrics, dists = self._get_aggregate_metrics_and_dists(timestamp or time())
  File "datadog/threadstats/base.py", line 348, in _get_aggregate_metrics_and_dists
    rolled_up_metrics = self._metric_aggregator.flush(flush_time)
  File "datadog/threadstats/metrics.py", line 173, in flush
    past_intervals = [i for i in self._metrics.keys() if i < interval]
  File "datadog/threadstats/metrics.py", line 173, in <listcomp>
    past_intervals = [i for i in self._metrics.keys() if i < interval]

[zippolyte]

This LGTM, thanks a lot for this @TheKevJames

It should fix #365

[edkorthof]

Thanks for the fix. But to clear -- this wasn't just unsafe in python3; in python2 it was also unsafe (but less likely to trigger); I provided a script which allows reproduction in python2.7.