feat: add vad_parameters to Predictor

[mazy06000]

What

This PR extends the predict() function to accept a new vad_parameters argument. The parameter is forwarded to model.transcribe(...), enabling custom control over VAD settings such as silence threshold and speech padding.

Why

• Improves flexibility for users seeking fine-grained voice activity detection.
• Empowers developers to reduce segment length and avoid repeated content by tuning VAD.
• Addresses community-requested enhancement; enhances overall user control.

How

• Updated predict() signature to include vad_parameters=None
• Passed vad_parameters through to model.transcribe(...) call
• Updated handler to parse vad_parameters from user payload
• Updated INPUT_VALIDATIONS schema to validate vad_parameters as a dict with a default of None.

Example usage

{
  "input": {
    "enable_vad": true,
    "vad_parameters": {
      "min_silence_duration_ms": 50,
      "speech_pad_ms": 20
    },
    ...
  }
}