Benchmark TDAG/CDAG generation in isolation

[fknorr]

This PR is similar to #62 in that it benchmarks various graph topologies, however, it only measures graph generation and serialization as they happen on the master node without actually passing commands on to workers. It can therefore be evaluated on a single node and timed more precisely, but does not benchmark job generation or execution at all.

• [task-graph] benchmarks measure TDAG generation.
• [command-graph] benchmarks measure TDAG + CDAG generation and serialization in the same thread
• [scheduler] benchmarks are like [command-graph], but use a dedicated scheduler thread similar to the runtime

I shamelessly copied the artificial graph topologies from @PeterTh 's #62 and added real-world graph examples from wave_sim as well as an (imaginary) Jacobi solver implementation.

Benchmarking the scheduler thread required some refactoring that allows re-using its std::thread in order to exclude the thread creation overhead from timings.

[fknorr]

I have simplified the implementation of benchmark tree graph generation and introduced more complex communication patterns. The correctness of the generated graphs can be verified through the [debug-graphs] test cases.

[psalz]

Thanks, LGTM! Given the precedent set by #100 this should also include updated benchmark results for ci/perf.