Improve Japanese TTS/STT UX and dependency handling

[TechNavii]

Context

MLX Audio’s Japanese TTS/STT UX had several rough edges: dependencies caused runtime crashes, the TTS language picker allowed unsupported combinations, the download button did not work time to time and transcript playback defaulted to a placeholder clip. This PR focuses on stabilising japanese so users can reliably synthesize and transcribe in Japanese.

Description

• Harden the backend dependency chain for Kokoro TTS (auto-install friendly requirements, configure fugashi/MeCab
  dictionaries, and guard language requests).
• Expand the TTS UI so language options reflect the selected model’s capabilities, automatically select valid voices,
  and generate real downloads.
• Simplify STT uploads to always auto-detect language, capture the uploaded audio alongside the transcript, and
  improve the transcript playback experience.

Changes in the codebase

• requirements.txt, mlx_audio/server.py, mlx_audio/tts/utils.py, README.md: add misaki extras (misaki[ja], misaki[zh],
  pyopenjtalk, fugashi[unidic-lite]), alias save_weights for older mlx_lm, bootstrap fugashi/MeCab env vars, and
  document the behavior.
• mlx_audio/ui/app/text-to-speech/page.tsx: enforce per-model language support (auto-detect only when supported),
  auto-select matching Kokoro voices, and implement a working audio download button.
• mlx_audio/ui/app/speech-to-text/page.tsx + [id]/page.tsx: remove the manual language dropdown, always send empty
  language for auto-detect STT, persist the uploaded audio data URL, feed it into the transcript player, and fix the
  playback timer.

Changes outside the codebase

• No infrastructure or third-party service changes beyond the new pip dependencies (misaki extras and fugashi’s
  UniDic-lite). The server now configures MeCab automatically; no manual OS-level install steps are required.

Additional information

• The TTS language guardrails automatically downgrade to the first supported language when switching between models
  (e.g., Marvis → English only).
• Transcript playback now uses the actual uploaded audio; durations come from metadata rather than static defaults.
• STT auto-detect keeps working with Whisper-family models without extra backend logic.

Checklist

• Tests added/updated
• Documentation updated
• Issue referenced (e.g., Closes Model type kokoro not supported #123)