perf(speech): replace speak settings polling with listener

[JustYannicc]

Summary

• Replaces the speak manager's 2s settings polling loop with one initial settings load plus the existing settings-updated preload event.
• Keeps the configured TTS model/voice and runtime speakOptions sync behavior in a small tested helper.
• Cleans up the settings listener on unmount.

Before / After Evidence

• Baseline instrumentation before product edits found idle getSettings calls at t=0, 2000, 4000, 6000ms, or 4 calls over a 6.5s idle window.
• After this change, the hook has one initial getSettings() call on mount and no settings polling interval. While idle, additional settings reads are 0; live changes arrive through onSettingsUpdated.
• This removes roughly 30 idle getSettings calls per minute per mounted speak manager, about 1,800 calls per hour in a long session.

Validation

• node scripts/test-speak-settings-listener.mjs passed.
• node --test 'scripts/test-*.mjs' passed: 28 passing, 1 skipped Electron-live test.
• node_modules/.bin/tsc -p tsconfig.main.json passed.
• node_modules/.bin/vite build passed.
• Codex LSP diagnostics on useSpeakManager.ts and speak-settings-sync.ts: no errors.
• node scripts/check-i18n.mjs exited 0; it reports two existing Italian missing keys unrelated to this change. No user-facing strings were added or changed.

Notes

• npm is intentionally disabled on this machine and its wrapper attempted a pnpm install that is blocked by pnpm's ignored-build policy, so validation used the underlying script commands directly from node_modules.

[JustYannicc]

Superseded by consolidated upstream PR #534 (perf(mac): reduce native polling and async IO stalls), which carries this change together with the related macOS/native/audio/camera async IO performance fixes. Closing this source PR so review can continue on #534.