feat(inference): add InferenceWebsocket

[chenghao-mou]

A new websocket helper that:

• provides a context manager for lifetime management
• recv loop that yields valid payloads, handles close detection internally via mark_closing()

Design decisions:

• It still leaves task orchestration and session creation to the caller
• It does not handle retry or stream class-level abstraction (RecognizeStream and InterruptionStreamBase already handle them differently in _main_task)

Usage:

async with InferenceWebSocket(
      session=http_session,
      base_url="https://agent-gateway.livekit.cloud/v1",
      path="/stt?model=deepgram/nova-3",
      api_key=api_key,
      api_secret=api_secret,
      timeout=10.0,
  ) as iws:
      await iws.send(session_create_data)
      async def send_coro():
          ...
      
     async def recv_coro():
          async for raw_data in iws.recv(str):
              ...

      tasks = [
          asyncio.create_task(...),
          asyncio.create_task(...),
      ]
      try:
          await asyncio.gather(*tasks)
      finally:
          await aio.cancel_and_wait(*tasks)