Langroid

fix: XMLToolMessage - parsing errors now return msg to LLM instead of crashing

fix: Improve XMLToolMessage formatting instructions sent to LLM

feat: add async streaming callbacks in addition to existing sync streaming callbacks

Mainly used with Chainlit currently. See:

- Relevant Pull Req 578
- [Chainlit callbacks](https://github.com/langroid/langroid/blob/main/langroid/agent/callbacks/chainlit.py)
- [Callbacks in `Agent`](https://github.com/langroid/langroid/blob/main/langroid/agent/base.py)
- [Callbacks in `LanguageModels`](https://github.com/langroid/langroid/blob/main/langroid/language_models/base.py)

fix: XMLToolMessage now handles complex fields like list, dict or nested Pydantic objects.

small fixes

- create_file() - allow append/overwrite/skip/error options
- `ChatAgentConfig.allow_multiple_tools` flag to control whether LLM can generate multiple tools at once (default True)

fixes:

- DocChatAgent improve check for collection missing in vecdb
- in ChainlitAgentCallback, set `llm.config.async_stream_quiet = False` to allow streaming output in chainlit