Validate benchmarks reliability

[crisidev]

Since #1230 we run a benchmark using wrk against origin/main and any pull request and calculate the deviation.

We need to validate if the benchmark is reliable when run by the Github action runner. If it is not reliable we can

• Run the benchmark on hardware we own
• Change the benchmark to count instruction instead of using time

[david-perez]

I've been keeping an eye on these. Usually the deviation does not go over 50%, but here's a PR in which we had 94% in one message, 16% in the next one #1368.

[david-perez]

Change the benchmark to count instruction instead of using time

Instruction count is is a useful metric for a compiler but not for what we're building. We could introduce severe performance degradation without noticeably increasing instruction count (think e.g. replacing a linear-time query string parsing function with a quadratic one).

[david-perez]

Closing since benchmarks were removed years ago because of unreliability, we'll use #214 as the tracking issue to re-add them.