v0.9.2: Stability, Request Handling & Voice Sorting

Curated summary: CHANGELOG.md

• fix(normalizer): read a stray comma before a unit as a decimal separator by @bfs2021 in #559
• fix(normalizer): space an unspaced ellipsis so short words aren't spelled out by @remsky in a6dcb9c
• fix(downloads): age temp files against the clock by @davidcoulson in #563
• fix(dev): return every segment from /dev/phonemize by @davidcoulson in #569
• fix(dev): validate the voice on /dev/generate_from_phonemes by @davidcoulson in #564
• fix(inference): retry a CUDA OOM once, before any audio is sent by @davidcoulson in #562
• fix(tts): fail the request when a chunk fails instead of dropping it by @davidcoulson in #571
• fix(model): serialise backend loading and refuse unload mid-generation by @davidcoulson in #572
• fix(inference): run model calls off the event loop by @davidcoulson in #566
• fix(voices): bound voice-mix files and caches by @davidcoulson in #565
• chore(lint): enable ruff ASYNC rules and run ruff in CI by @remsky in #574
• fix(api): cap request body size and bound dev text fields by @davidcoulson in #567
• feat(web): add A to Z sort toggle to the voice list, closes #542 by @remsky in #575

New Contributors

• @davidcoulson made their first contribution in #563

Full Changelog: v0.9.1...v0.9.2