feat(speech): add async and WebSocket TTS operations

[octo-patch]

Reason: Add the missing async and WebSocket TTS operations to the speech CLI and SDK.

Changes

• Add mmx speech async — create a long-form async TTS task via POST /v1/t2a_async_v2 (textToAudioAsyncCreate). With --wait it polls until completion and downloads the audio.
• Add mmx speech task get — query async TTS task status via GET /v1/query/t2a_async_query_v2 (textToAudioAsyncQuery).
• Add mmx speech websocket — synchronous streaming TTS over WSS /ws/v1/t2a_v2 (textToAudioWebSocket). Supports --out to save the audio and --stream to pipe raw audio chunks to stdout.
• Add SpeechSDK.createAsync, SpeechSDK.queryAsync, SpeechSDK.downloadAsyncFile, and SpeechSDK.synthesizeWebSocket.
• Add tests for the SDK async/WebSocket methods and the new CLI commands; document the commands in README, SDK.md, and docs/cli-design.md.

Checks

• bun run typecheck — pass
• bun run lint — pass
• bun test — 460 passing
• bun run build — pass

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

[NianJiuZst]

Thanks for adding the missing async and WebSocket TTS surfaces. I do not think this is ready to merge yet because the async path diverges from the current official API contract in three blocking places:

1. The speech async --wait download is unauthenticated.
   src/commands/speech/async.ts passes the protected /v1/files/retrieve_content URL to the generic downloadFile helper, which performs a bare fetch without an Authorization header. A successful async task will therefore fail at the final download step with 401. The official Retrieve Content API requires Bearer authentication:
   https://platform.minimax.io/docs/api-reference/file-management-retrieve-content

2. The async request uses the wrong sample-rate field.
   The new async request type reuses the synchronous HTTP TTS audio_setting shape and sends sample_rate. The async endpoint expects audio_sample_rate, so the new --sample-rate option is not wired to the documented request contract. See the official async-create request example:
   https://platform.minimax.io/docs/api-reference/speech-t2a-async-create

3. The advertised 1M-character path is not implemented.
   --text-file currently reads the local file and sends its contents as text. The official API limits direct text input to 50,000 characters; the larger long-form path requires uploading a file with purpose t2a_async_input and passing text_file_id. Neither the CLI nor SpeechAsyncRequest currently supports that path:
   https://platform.minimax.io/docs/api-reference/speech-t2a-async-create

I verified this on head ed939d1. Typecheck, lint, build, the 11 new focused tests, and GitHub CI pass, but the tests only exercise permissive mocks. A request-capture reproduction on the actual command path observed audio_setting.sample_rate and a null download authorization header.

Please fix these contracts and add regression coverage that asserts audio_sample_rate, authenticated content retrieval, and the text_file_id long-text flow before merging.