Exclude aggregates with anomalously low sample counts from alarm evaluation

Currently we have a potential scenario whereby an autoscaling alarm defined over an instance group, with the evaluation window & cooldown policy both on the short side, may fire prematurely if an outlier instance is one of just a small number reported as yet for the current period.

We should detect the case where the sample count is anomalously low for the most recent datapoints, based on say a variance of at least one standard deviation from the observed norm. In that case, the less reliable datapoint should be excluded from consideration for alarming.