Adding the ability to run restore_cache on a step failure would allow for saving time on potentially expensive post-failure actions.

This could help with cleanup steps or aggregation steps.