Investigate Examples performance

[tybug]

While investigating #4014, we improved on some performance hits we took during the IR migration (#3921). A remaining holdout is computing _mutator_groups, accounting for 15% of runtime in #4014 (comment), in part composed of _ir_starts_and_ends (7% total runtime).

This seems high to me. We should see if we can track equivalent information inside ConjectureData in a more performant way, for either ir bounds, or mutator groups, or both — or neither, and decide no action is needed.