Add custom CSV and Markdown reporters for benchmark results

[psalz]

This adds a custom CSV reporter for producing machine-readable benchmarking results. You can see an example output in the updated results for gpuc1 (I noticed some of the numbers diverge quite wildly from the previous results, not sure how reliable those are - cc @PeterTh).

To use the reporter, simply call the benchmarks with --reporter celerity-benchmark-csv.

[fknorr]

Since the output has is for machine use anyway, IMO we should just export individual samples and leave all the statistics to the postprocessing script (e.g. scipy + matplotlib). That would be strictly more general.

[psalz]

I've changed the CSV reporter to also include the raw samples. I've also introduced a new Markdown reporter that can be used to produce human-readable results alongside the CSV. Simply run with ./test/benchmarks --reporter celerity-benchmark-csv::results.csv --reporter celerity-benchmark-md::results.md to generate both simultaneously.

[fknorr]

Thanks! Most of this looks great, however for the Markdown output, we technically need to escape all the special characters so test / benchmark names aren't interpreted as formatting. For example, the CDAG benchmarks have an underscore as in wave_sim topology.

This is less about nitpicking and more towards the question whether we actually need proper Markdown output, or if a .txt file with an ASCII table (which essentially looks the same source-wise) would be enough. Viewing the Markdown file "raw" makes it much easier to parse for my eyes, since the numbers are properly aligned and spacing is not as wasteful (on GitHub's renderer).

Edit: I just realized GitHub has a (searchable) CSV renderer, which is kind of neat, but it unfortunately doesn't right-align numbers.

[PeterTh]

It looks like adding and removing dependencies and checking dependencies got much slower compared to the currently-checked-in results for the low-dep-count tests. It's still very fast, but don't really have an explanation for that.

[psalz]

This is less about nitpicking and more towards the question whether we actually need proper Markdown output, or if a .txt file with an ASCII table (which essentially looks the same source-wise) would be enough. Viewing the Markdown file "raw" makes it much easier to parse for my eyes, since the numbers are properly aligned and spacing is not as wasteful (on GitHub's renderer).

As discussed during yesterday's meeting, there is value in GitHub's Markdown diff, so we'll keep it. I've added rudimentary escaping.

[fknorr]

The markdown report in ci/perf/gpuc1_bench.md has not been updated since escaping was added. Otherwise LGTM!

[psalz]

The markdown report in ci/perf/gpuc1_bench.md has not been updated since escaping was added. Otherwise LGTM!

I've manually added the escaping to the relevant benchmark names; rebased onto master.