Switch otf2 metric mode from accumulated_start to accumulated_last

[cvonelm]

Currently, we are using accumulated_start for perf metrics. In this mode, the metric value we write is the sum of all metric readouts since the beginning of measurements.

Alternatively, afaik, we could use accumulated_last, where we write the individual metric readouts instead.

This is not the highest priority of issue but to my knowledge it seems that the accumulation we are doing in lo2s is unnecessary.

And additionally, with accumulation we always run the risk that issues such as #267 mess up the accumulation variable by adding values such as NaN or infinity.

[tilsche]

IIRC this was done primarily because of long-standing bug(s) in Vampir. We have to carefully check if those are properly fixed.