Add an acceptance test that triggers a config reload and verifies
that no early notification is occurring.

One of the challenges is which time to use to check for a previous
notification. The nflog captures about the time all notifications
were sent. That conflicts with the ag.next timer that get's reset
before the ag is being flushed. Delays and retries can make these
two point in time be different enough for the integration tests to
fail.

I considered the following ways to fix it:

  1.) Modify the nflog.proto to capture the flush time in addition
      to the successful notification time.
  2.) In addition to hasFlushed capture the last flush time and pass
      it to the context like we do for Now.
  3.) Set hashFlushed and reset the timer after the flush has been
      done.

I started with prometheus#3 as it seemeded to have the fewest downsides with
things like drift. Based on comments this is no prometheus#1.

needsUpdate is based on:
prometheus#3074 (comment)

Signed-off-by: Holger Hans Peter Freyther <holger@freyther.de>

When creating an AggregationGroup consult the Stage about the
next expected time to `Exec` and use it to set the initial timer.

Update the acceptance test to match the new behavior and add a
test to receive the updated group at group_interval timeout.

Signed-off-by: Holger Hans Peter Freyther <holger@freyther.de>

Use the dispatch time instead of the last successful notification
for the repeat interval. This avoids a drift and also keeps the
same time during notifications.

Signed-off-by: Holger Hans Peter Freyther <holger@freyther.de>