feat(types): add Audio class and audio field to Message for multimodal models

[Ghraven]

Summary

Adds a dedicated Audio class and audio field to Message, mirroring the existing Image pattern.

Closes #650

Motivation

Currently, audio data must be passed via the images key, which is confusing and blocks future models that support both images and audio simultaneously. This PR adds a first-class audio field so callers can pass audio data cleanly:

ollama.chat(
    model="gemma4:e2b",
    messages=[{
        "role": "user",
        "content": "Transcribe this",
        "audio": ["recording.wav"],   # clear, not crammed into images
    }]
)

Changes

ollama/_types.py

• Added Audio(BaseModel) class after Image, with identical serialisation logic:

  • Path / bytes → base64-encodes the data
  • str path that exists on disk → base64-encodes the file
  • str with a known audio extension that doesn't exist → raises ValueError with a clear message
  • str that looks like existing base64 → passes through
  • Unknown string → raises ValueError
  • Supported extension check covers: mp3, mp4, mpeg, mpga, m4a, ogg, wav, webm

• Added audio: Optional[Sequence[Audio]] = None field to Message (after images), with the same docstring style as images.

Compatibility

• No breaking changes — audio is optional and defaults to None
• The serialisation behaviour is consistent with Image, so the wire format is already what the Ollama server expects (raw base64)