benchmark: refactor to be usable on CI for releases

[bzz]

Current benchmark takes ~15h on a modern laptop to finish.

It may be either too detailed, as final results are averaged anyway, or may be using too much data (all samples from linguist right now).

The goal is to have a benchmark that:

• agrees with current published data
• short: takes >1min to finish (part of local dev workflow)
• long: takes ~10 min max to finish (part of every release workflow)
• long: capable of runing with origuma, to highlight the performance improvements by switching regexp engine