engine: simplify in-memory aggregation logic. add instrumentation.

[pyr]

This commit brings in a number of changes:

• Handle input drift globally. A new component is introduced.
  to facilitate testing, the wall-clock is abstrated into
  its own component as well which the drift component depends
  on.
• Simplify the way the engine holds on to values in memory.
  Values are stored in an atom and updates trigger a
  swap! on the atom, yielding snapshots.
• Move queueing mechanism into a proper component
• Allow per-queue capacity and pool-size parameters. A good
  rule of thumb is to make room for twice your average input
  size, which gives the system some time to cope with bursts.
• Instrument critical operations, such as taking snapshots
  these are available as JMX by default. Configuration support
  will be provided to handle reporting destinations.

[pyr]

This takes care of #121 which is able to handle large loads without suffering at -Xmx512m additional tuning will make it possible to go further.

It should be noted that the default in-memory store doesn't cope too well with large payloads and will be adapted. In the mean-time large clusters are encouraged to rely on the elasticsearch index.