Blin runs should be resumable

[coke]

For long runnings tasks like Blin, we should be able to lose only the work in progress on a given run, and resume the work in progress instead of starting a new run.

This will come in handy for running on (e.g.) Azure spot VMs which may be shut down at any time, but allow for a cheaper cost when they are running.