tests: try to solve false negative on execution time benchmark

- new tunning of the limit threshold (Fail on 4ms RunModel with a gap of 0.25
- move benchmark test from check to a specific job and only for patched version

Refs #136