Using the absolute path fixes the energy measurements because perf was
called in $HOME with a relative path inside emper-io-eval which did
break.
And using only a single eval.py call means we only hav one desc.yml file
containing everything for that run.

pipe performs generally better than vanilla for network io.

* introduce a Makefile for the generated plots
* allow throughput and latency plots to be generated individually
* introduce new best plot including all runtimes and waitfd
* exclude the new computation 0us-100ms-0.5

* add output file support.
* create multiple plots at once.
* use less extreme percentiles

Generate and evaluate '0us-10ms-0.5' emper flavors.

Energy measurement tooling does not seem stable yet.

I expect the 99.999 percentile mostly to include extreme latencies
introduced by the experiment setup.
Nearly all flavors show some latencies around ~13500 and ~26000 ms.

Replace all occurrences of computation* with vanilla*.

This is needed for the latency evaluation.