deflake imageGenQuota bus tests by draining the handler instead of sleeping 20ms

[atomantic]

Summary

• imageGenQuota's bus handlers are fire-and-forget (an EventEmitter cannot await its listeners), so the suite slept a fixed 20ms before asserting the persisted ledger. The recorder's write is real disk I/O, and on Windows atomicWrite additionally sleeps between rename retries when the destination is briefly locked — so the write can outlast any constant a test picks. That is the flake: subscribes only once across repeated boots failed with expected 'No renders · 24h' to be '1 render · 24h' on the Windows runner and passed on a same-SHA re-run.
• The hook now keeps a single promise tail over every outcome its handlers dispatched, and exposes __drainImageGenQuotaHookForTests() so the suite awaits the actual write instead of guessing at a duration. The drain loops until quiescent, so work a settling handler chains on is drained too.
• Production behaviour is unchanged: dispatch still happens immediately and is still fire-and-forget; the tail only records what was dispatched. The handler's existing .catch means the tail can never reject or go unhandled.
• Scope: this fixes only the named case's timing dependency. It does not touch the unrelated Windows-CI failures that have been red on main for other reasons.

Test plan

• npx vitest run services/imageGenQuota.test.js — 32 passed, run 10× consecutively, green every time.
• Not vacuous — bypass probe. Temporarily removed the if (subscribed) return; guard in initImageGenQuotaHook (breaking the "subscribe once" invariant) and confirmed the test goes red with expected '2 renders · 24h' to be '1 render · 24h'. Restored.
• Reproduced the reported failure. Injected a 200ms artificial delay into recordImageGenOutcome: with the new drain all 32 tests still pass; with the old 20ms sleep restored, the suite fails with exactly the CI message expected 'No renders · 24h' to be '1 render · 24h'. Both probes removed before committing.
• Full server suite (cd server && npm test) run twice: 32,714 / 32,752 passing. The only failures were services/imageTo3d/trellis2NormalBake.test.js and routes/settings.secretsStrip.test.js hitting the known 10s timeout under full-suite load — a different pair each run, and both pass in isolation. Neither imports imageGenQuota.

Closes #4788