Improve aggregation

[untitaker]

The aggregator has a few issues:

• durations/distributions are not aggregated at all, we could add reservoir sampling at the very least
• aggregator checks the system time too often, causing overhead. instead we should take over @Swatinem's design where there's a background thread that sleeps + flushes out the state
• ingest team reported that statsdproxy's aggregator sometimes creates wrong numbers (to be investigated)