Explore possibility of time-based tests for trappy

[joelagnel]

Recently lot of speed up related patches have posted. It would be good to be able to test and make sure these improvements aren't regressed in the future. One concern is that its hardware dependent and is hard to measure. This issue is created to discuss this.
I think we can still come up with a worst-case upper bound experimentally and set that as the expectation. If we are threading trappy in the future, then we can perhaps limit the number of threads so that the number of cores that the test is run on doesn't change or artificially make a failure result as passed (as a false-positive). (CC @derkling)