The wrong `size_class` was used when fetching stack traces from empty
buckets. The `size_class` would always be the maximum value after
exhausting the search of active buckets rather than the actual
`size_class` of the allocation.