feat!: replace qwen-tts with qwen3-tts-vc-2026-01-22

[LingXuanYin]

Summary

• qwen-tts is retiring on DashScope (2026-10-10). Replaces it with qwen3-tts-vc-2026-01-22, the long-term-supported voice-clone model on the worker side (background/tasks/qwen_tts_actor.py).
• qwen3-tts-vc-2026-01-22 is clone-only (no voice-design mode exists for this DashScope model family), so it gets its own adapter validate/buildPayload path instead of joining QWEN_MODELS.
• qwen-audio-3.0-tts-plus/-flash are untouched — already live/shipped, unrelated to this migration.
• cosyvoice-v3.5 is intentionally not added: it never shipped (only existed on the now-closed feat!: replace qwen-tts with cosyvoice-v3.5-plus/flash #48), so there is nothing to keep exposed here. The worker independently keeps serving it for any caller with a pinned target_model.

Replaces

Closes #48, which targeted cosyvoice-v3.5 instead — superseded by the qwen3-tts-vc direction above.

Test plan

• pnpm test (vitest run) — 214 passed
• pnpm typecheck (tsc --noEmit) — clean
• models/qwen3-tts-vc-2026-01-22.yaml generated via neta-generation models export and verified to round-trip against the runtime declaration

[LingXuanYin]

Reverting this: further verification against Alibaba's own retirement notices shows qwen3-tts-vc-2026-01-22 is ALSO scheduled to retire on 2026-10-10 (same day as qwen-tts), with cosyvoice-v3.5-plus as the officially documented replacement:

• https://www.alibabacloud.com/en/notice/model_studio_notice_of_retirement_for_selected_legacy_models_7d9 (explicit table row: qwen3-tts-vc-2026-01-22 → cosyvoice-v3.5-plus)
• https://www.alibabacloud.com/en/notice/model_studio_notice_on_the_postponed_retirement_of_selected_legacy_models_7d0 (the qwen-voice-enrollment model used to create a voice for the qwen3-tts-vc family is also on the retirement list)

Corrected direction: cosyvoice-v3.5-plus/-flash share the exact same wire shape as qwen-tts (same voice-enrollment creation model, same SpeechSynthesizer endpoint/fields) -- so this SDK's qwen-tts declaration needs no change at all; the model swap happens entirely downstream in background/new-api. Closing this PR since there is no SDK-level change to make.