Consider lighter-weight counters for optimized code

> A design thought: I wonder if it would make sense to have a mode where we *only* have the `incorrect` counters? The reason being that it might be a nice property to avoid overhead in the high-prediction-accuracy case -- and we might be able to come up with some other reasonable denominator (like total runtime) to judge whether `incorrect` crosses a badness threshold in a given system.

_Originally posted by @cfallin in https://github.com/fitzgen/winliner/pull/2#discussion_r1333585336_

Start with the linked discussion in pull request #2 and trace where the existing counters are defined and used. Determine whether an incorrect-only mode and an alternative denominator can be specified without losing useful threshold information; done means the design, overhead trade-offs, and badness threshold are agreed.

Written by the indexing model from the issue text.