Vouch request: fix for SQLite "database is locked" (#3679)

[fitzyracing1]

Hi, I'm @fitzyracing1. I'd like to contribute a fix for #3679, the intermittent (code: 5) database is locked error in persistence/sqlite.rs.

kvnloo handed the issue to me on Oct 3, and I posted my stress-test findings there on Oct 7. In short, no single path holds the write lock for more than 5s. Instead, a waiting writer gets starved by a steady stream of short writes. The same error has since shown up in the #4408 test on the Windows MSVC lanes and on a real gateway with two clients (#4591).

The fix sets the 5s busy_timeout explicitly and adds an in-process first-come, first-served write queue to SqliteStore. It covers every write path, including the image transaction in sandbox_image.rs. Reads and the startup migration aren't gated. I tested it with 50 ms added to every disk sync. The #4408 concurrent-writes test passed 30/30 runs with the fix and 0/30 on main. On a normal disk, both passed 100/100. The full openshell-server lib suite passes, and fmt and clippy are clean. I haven't run the E2E or Windows tests.

The change is ready as #5102, which was auto-closed pending a vouch. I'm happy to adjust the approach based on maintainer feedback.