Compare perf with base branch on CI

Currently, for each PR CI measures performance of the library and publishes a comment with results (see any recent PR as an example). It's a lot of numbers and it's difficult to see if there's any regression. 

CI should also benchmark the lib before the PR, and compare the results. It should help us see if there's any performance regression/improvement introduced by the PR