fix metric calculation caused by mismatched pair comparisons

[kleinercubs]

The evaluation results must be saved in the order of their keys (problem_id). Otherwise, when some eval results are missing, the speedup computed with geometric_mean_speed_ratio_correct_only would use mismatched pairs and produce incorrect results.

[simonguozirui]

Thanks for catching that @kleinercubs.

@pythonomar22 let's look into this as we are revamping metrics calculations and summary statistics.