fix: refactor DoubaoBot to use AsyncArk with proper async streaming

[octo-patch]

Closes #594

Problem

The DoubaoBot implementation has several issues introduced when it was
migrated to the Volcengine Ark V3 SDK:

1. Blocks the event loop: Both ask and ask_stream call the
   synchronous Ark client inside async def methods without await,
   blocking asyncio while waiting for network I/O.

2. Streaming loop uses a sync iterator: ask_stream used for chunk in … instead of async for chunk in …, which prevents asyncio from
   yielding control between chunks.

3. Missing split_sentences: Other bots (chatgptapi, ppio, jiekou)
   pass the raw LLM token stream through split_sentences before
   yielding to the TTS layer, producing natural sentence-boundary
   chunks. DoubaoBot was yielding raw tokens, resulting in choppy or
   broken TTS output.

4. No guard for empty choices: A chunk with an empty choices
   list would raise IndexError. chatgptapi_bot guards against this
   with if not event.choices: continue.

5. Deprecated default model: skylark-chat is an old Ark V2 model
   name. The current Ark V3 equivalent is doubao-pro-32k (users can
   still override via gpt_options).

6. ask returned None on error instead of "", violating the
   BaseBot abstract method signature.

Solution

• Swap Ark → AsyncArk and add await to chat.completions.create
• Iterate the streaming response with async for
• Wire split_sentences into ask_stream, matching ppio/chatgptapi pattern
• Add if not chunk.choices: continue and if content is None: continue guards
• Change default model to doubao-pro-32k
• Return "" instead of None from ask on error

Testing

Changes follow the exact patterns already used by ppio_bot.py and
chatgptapi_bot.py, which are known to work correctly in production.