Improve benchmarking infra

[lorenzwalthert]

• The variance in the timing estimates are probably > 1%. A PR and a base reference where no effect on speed is expected (because no such code was touched, Depreciate env var save_after_styling #631), gave a ~2% difference.
• add time-line plot.
• artifact should have name of branch when uploaded to later easily compare different artifact.
• absolute median timing should be posted in PR for easy reference.
• Need an easy way to show median is x% worse or better than base branch. Currently, we need to estimate from hand and first guess the median. Could be in the plot or posted as comment.
• Need additional benchmark. Now we have without cache and with cache. The usefulness of with_cache is limited because it is so close to 0 (since it's always full text cached already). We don't have with cache activated and empty cache. This would be helpful for understanding the impact of Speed improvement in caching  #679 for example, or any improvement that relates to the cache. The interpretation of it would include the time required for git reset and can hence not be taken as absolute time to style code, but only in comparison with the base branch (where git reset also occurs).

[lorenzwalthert]

Closed with #698.