Loose-mode evaluation is not implemented in instruction_following #1160

#1158 introduces ifeval, but does not report all the same metrics as Skills:

Skills' IFMetrics reports four scores (prompt_strict_accuracy, instruction_strict_accuracy, prompt_loose_accuracy, instruction_loose_accuracy) plus their average; the existing Gym server reproduces strict mode only. Adding the eight-perturbation loose check to verify() and a custom compute_metrics() would fully reproduce Skills' breakdown.