Simplify pending work tracking

[mkeeter]

Alan and I were talking in person about pending work tracking in the Downstairs, and I found some room for simplification:

• Instead of tracking pending work and number of unmet dependencies in separate hashmaps, put them into a single struct PendingJob
• Instead of using HashMap<JobId, IOop> then cloning + sorting the keys to make a Vec<JobId>, just use a VecDeque

This are relatively small cleanups, but may have a small impact on Downstairs RAM usage.

There are two behavior changes, both of which should be innocuous:

• We now check whether jobs are ready in the order that they're received, instead of sorting all pending jobs by JobId then checking whether they're ready. This is fine, because readiness is determined by dependencies.
• The show_work debug function no longer sorts job IDs when printing pending work