Change significance to be determined by IQR fencing

[rylev]

This changes how we define (and subsequently implement in code) a "significant" test result to a more formal and less arbitrary mechanism (described below). Additionally, the documentation is updated to reflect this change.

Before
Until we've used a simple threshold of either 0.2% change for non-"dodgy" test cases (i.e., test cases which we've determined to not have some sort of historical noise) and 0.8% for "dodgy" test cases.

After
Significance is defined as being an outlier when compared with historical data. We use interquartile range fencing to determine whether a given result is an outlier.

IQR fencing uses this formula:

interquartile_range = Q3 - Q1
result > Q3 + (interquartile_range * 1.5)

[rylev]

I'd like to hold off on merging this for now since it makes a lot more comparisons show up as "definitely relevant". We should address that first.