[Proposal] Failure avoidance for Current_docker.pull

[kit-ty-kate]

Currently when something fails during a Current_docker.pull its output will stay a failure until the next scheduled pull. For this particular command we probably don't want that. Here is my proposal to improve this case:

1. If a pull fails, then just take its previous value.
2. If there was no previous value, try again a minute later, then every hour if it still fails.
3. Unless the service just started and the failed previous value is to be taken from the disk cache, in which case, try right away, then every hour if it still fails.

[talex5]

More generally, we need an option to latch the old successful result of a step when a new build fails. We'll also need a way to report a metric for such hidden failures.

Retrying failed jobs on a schedule is also important, but should probably be a separate issue.