Explaining the variation in the attained power of a stepped-wedge trial with unequal cluster sizes