Mempool benchmark is flaky

[nfrisby]

The CI on three PRs that do not change any code (309, 311, 312) failed due to the mempool benchmark.

[amesgen]

I have seen occasional (definitely flaky) failures of the mempool benchmark since their inception (might be related to higher GitHub Actions infra load?). We should definitely fix this to both remove the scary red marks and make the benchmarks more indicative, such that they could even be marked as "Required" eventually (they are not currently).

Potential approaches:

• Increase the tolerance even further.
• Experiment with custom hosted runners from IOG, in case they have more predictable/consistent performance.

[jorisdral]

Another potential improvement, which could serve as an intermediate solution:

• Upload the built benchmarking executable as an artifact/cache and have a separate job/workflow that runs the bechmarks. We can then mark the build/test job as required, leave the benchmark job as not required, and we save time spent running CI because we're not rebuilding/retesting every time the benchmark step failed.