fix(audio): resolve model alias lookup for STT/TTS/STS endpoints

[haokaiyang]

• Fixes audio endpoints not resolving model aliases before loading engines.
• Adds alias resolution in:
  - /v1/audio/transcriptions
  - /v1/audio/speech
  - /v1/audio/process
• Introduces _get_settings_manager() for lazy access to server settings (keeps import pattern consistent with _get_engine_pool).
• Uses local model variable in TTS route instead of mutating request.model, improving handler-side consistency and reducing side effects.

[haokaiyang]

Fixes #489