fix(stt): STT compatibility fixes for Groq Whisper and AionUI web frontend

[starm2010]

Problem

When using AionUI web frontend with Groq Whisper as the STT provider, speech-to-text fails with 400/502 errors. There are four distinct issues:

1. Preference key mismatch: The AionUI web frontend stores STT config under the key tools.speechToText in client_preferences, but the backend queries only speechToText. Since get_preferences() filters with WHERE key IN (...), the config row is never found, causing STT_DISABLED even when the user has configured STT.

2. Multipart field name mismatch: The web frontend sends audio as FormData field audio (via formData.append('audio', blob, filename)), but the backend expects field name file. This causes a 400 Bad Request: missing 'file' field error. The frontend also embeds the filename in the blob's Content-Disposition header rather than as a separate multipart field.

3. Double /v1 in Groq URL: When base_url includes /v1 (e.g. https://api.groq.com/openai/v1), the code appends /v1/audio/transcriptions, producing https://api.groq.com/openai/v1/v1/audio/transcriptions → 502 Bad Gateway.

4. Language code and MIME type incompatibility: The browser sends languageHint: "en-US" but Groq Whisper only accepts ISO 639-1 base codes (e.g. en). Similarly, the browser sends audio/webm;codecs=opus as the MIME type, but reqwest::mime_str() requires clean MIME types without codec parameters.

Solution

1. Key mismatch: Query both speechToText and tools.speechToText in get_preferences(). Use prefs.get("tools.speechToText").or_else(|| prefs.get("speechToText")) for backward compatibility.

2. Multipart field: Accept both "file" and "audio" field names. Parse filename from the Content-Disposition header when the frontend sends it as part of the blob rather than a separate field.

3. Double /v1: Add .trim_end_matches("/v1") after .trim_end_matches('/') when constructing the base URL, before appending /v1/audio/transcriptions.

4. Language normalization: Strip region codes via lang.split('-').next() (e.g. en-US → en). User-configured language in settings now takes precedence over browser languageHint, so users can override the browser locale for transcription.

5. MIME normalization: Strip codec parameters via mime_type.split(';').next().trim() (e.g. audio/webm;codecs=opus → audio/webm).

Testing

• Tested with AionUI v2.1.10 web frontend speaking into the microphone
• Verified Groq Whisper returns correct transcriptions in both English and Spanish
• Verified curl tests directly against aioncore /api/stt endpoint return 200 OK
• Verified setting language: "es" in STT settings correctly overrides browser locale
• Confirmed no debug/temporary code in the final commit

Checklist

• Changes are minimal and focused on the STT compatibility issues
• No debug/temporary code included
• Backward compatible — both old (speechToText) and new (tools.speechToText) keys work
• Both "file" and "audio" multipart field names accepted
• No changes to error.rs — upstream already has the correct error model

---

Fixes #373