feat: add configurable OpenAI-compatible speech to text

[LauraGPT]

Summary

• add a browser-side multipart transcription client for any OpenAI-compatible /v1/audio/transcriptions service
• add persisted Settings fields for a FunASR/SenseVoice-compatible base URL, model, and optional API key
• add a toolbar audio-file action that appends returned text to the current draft without sending it
• document the local-service boundary; NextChat does not bundle a Python runtime or model weights

Verification

• corepack yarn test:ci test/stt.test.ts (3 passed)
• corepack yarn build (passed; existing lint plugin warning on app/constant.ts remains non-fatal)

Related to #6803. Keep that issue open for maintainer and reporter verification after merge.

[LauraGPT]

Exact head 9f6c5c77731c23615c3d17183b2bf8fc14f42875 has been revalidated locally:

• corepack yarn test:ci test/stt.test.ts (3 passed)
• corepack yarn build (successful)
• git diff --check

The Next build completes with existing non-fatal warnings from optional bufferutil/utf-8-validate, the lint plugin on app/constant.ts, and browser-only localStorage during static generation. The feature remains opt-in, sends multipart audio directly to the configured OpenAI-compatible endpoint, and does not bundle Python or model weights. Ready for maintainer merge; related issue #6803 remains open for user confirmation.

[LauraGPT]

Closing this duplicate in favor of #6860. The retained implementation provides the same configurable OpenAI-compatible multipart transcription flow while keeping web requests behind the existing same-origin proxy and using direct local calls only in desktop mode. That matches the original request's credential-boundary requirement and avoids a browser CORS requirement.

The original request in #6803 remains open for maintainer review; no user request is being closed here.