[FEATURE]: support xAI Grok TTS and STT in the AI speech/transcription routes #56041

Description

The @opencode/ai package ships speech generation and transcription routes for six providers (OpenAI, ElevenLabs, Cartesia, Deepgram, Google, AssemblyAI), but xAI's Grok Speech APIs have no route yet. The chat side already supports xAI, and the grok-voice-* realtime engine is in flight in #55951.

xAI documents both speech APIs as first-party REST endpoints:

• Text to Speech: POST https://api.x.ai/v1/tts — JSON in, raw audio out. Up to 60,000 characters, expressive voices (voice_id, default eve), language required (BCP-47 or auto), mp3/wav/pcm/mulaw/alaw output, speed 0.7–1.5, speech tags and pronunciation replacements.
• Speech to Text: POST https://api.x.ai/v1/stt — multipart with file last (or url), model grok-voice-transcribe-2.0, word-level timestamps always returned, optional diarize, repeated keyterm, inverse-text-normalization via format=true, up to 500 MB / 8 channels.

Docs: https://docs.x.ai/developers/model-capabilities/audio/text-to-speech and https://docs.x.ai/developers/model-capabilities/audio/speech-to-text

Proposed scope

• xai-speech: lower SpeechRequest to voice_id, language (defaulting to auto), output_format.codec, and speed; reject instructions, timestamp requests, and codecs xAI does not produce; describe the returned audio from the codec actually sent.
• xai-transcription: lower to the multipart form with file last (or a url field for URL sources); support diarize, word-timestamp answers, and repeated keyterm through provider options; map words/speakers, language, duration, and seconds usage.
• Tests: lowering and rejection coverage, plus replayed cassettes for both routes.

I have a working implementation and will open a PR.