ref(utils): Small SafeRolloutComparator refactors

[lobsterkatie]

This makes a few small changes to the SafeRolloutComparator code.

• For the initial round of span-first detector testing, only the comparison part of the comparator is needed. This therefore splits the check part of check_and_choose off into its own compare method.

• In cases where only the comparison is being used, it's not accurate to say that either the experimental or control data is being used. Therefore, "both" and "neither" have been added as allowable source_of_truth tag values, to account for (as the names would suggest) both results being used or both results being discarded.

• Finally, now that it's part of the compare method rather than the check_and_choose method, for new comparators the SafeRolloutComparator.check_and_choose metric has been renamed to SafeRolloutComparator.compare. (The one existing comparator now has a flag which preserves the old behavior, so as not to break existing dashboards.)

h/t Claude for his help with parts of these changes