Stablize compilation time measurement in benchmarks

[antiagainst]

Right now we measure the compilation time for compiling various models. Though the numbers reported can vary a lot. For example, this one ranges from 60s to 140s; similarly for CPU. With such a level of noisiness we cannot draw any meaningful information out of it and it's actually a distraction. We'd need to figure out ways to make the numbers more stable.

[pzread]

Unassigned myself as I'm not working on this currently