FileSystem appears to be threadsafe, but could probably use some more
investigation to ensure that we're not inadvertently causing a data race.

Add async overall and per-step JobSet queues to Pi::App for startup code.
Load steps now wait for async jobs to complete before progressing.
The loading screen now waits for all async jobs to transition to menu.

This takes the FaceParts init step (with prewarmed fs cache) to ~220ms from 440ms.

Run FacePart initialization on the background queue (not needed by further steps).

Remove further unnecessary profiling points.
This cleans up ~1/4th of the profile events during star system cache generation.

Instead of finishing one job per thread per frame, now copy job pointers to an on-thread array and run their finalizers in one pass.

This means we can queue potentially 1000s of jobs in a single frame / load step without then needing to finalize them over the next 250 frames.

Potentially investigate an upper limit on the number of finalized jobs; e.g. 100/thread/frame to avoid frame hitches.