Review the variability of the sharding benchmarks

This task is a placeholder to discuss about what to do with the sharding benchmarks. Some context: recently some safety checks were introduced to the benchmarks to verify that we don't add performance drops of more or equal than a 20%. The problem seems to be that our benchmarks have a huge variability and because of that we are generating a bit of redness on EVG.
It is interesting to point out that since these checks were introduced we have not managed to get a green execution of these benchmarks.