Properly Aggregate String Attribute Data

[michaelmckinsey1]

The following is an example from RAJAPerf for an inclusive metric, the blue bars are for a profile without string attributes and the orange bars are for a profile with the Complexity string attribute, added in llnl/RAJAPerf#471:

Notice that the orange bars should add up in the parent kernel Polybench to approximately 33.5, but are nowhere close. This is because the orange value shown for Polybench is showing the metric for only the O(N^(3/2) kernels, which if we look at the below example is correct:

This is misleading, first because you would expect all of the children kernels to add up to the parent kernel, and secondly because the Complexity attribute only appears when using from_caliperreader(..., string_attributes=["Complexity"]).

---

Proposed solution is to aggregate the parent kernel values (e.g. 22+11) during Hatchet/Thicket initialization.