Langroid

feat: `Agent` - add true async methods for `agent_response_async` and `user_response_async`

Previously these methods were simply calling the sync version, which limits their usefulness where we need truly async behavior.
This PR creates truly async versions of these.

Thanks alexagr !

feat(minor): Enhanced support for strict-decoding, with configurable options

feat: Support strict/guaranteed tools and structured JSON outputs

Leverages grammar-constrained decoding provided by OpenAI API as well as local LLM
inference providers such as `llama.cpp` and `vLLM`. In Langroid we do not directly specify the grammar;
instead we specify the desired output format as a type (simple type or derived from Pydantic BaseModel, or `ToolMessage`),
and we rely on the API to (implicitly) convert this format spec into a grammar, and guarantee
that the LLM output adheres to the specified format type.

- [docs](https://langroid.github.io/langroid/notes/structured-output/)
- [test_structured_output.py](https://github.com/langroid/langroid/blob/main/tests/main/test_structured_output.py)
- [test_tool_messages.py](https://github.com/langroid/langroid/blob/main/tests/main/test_tool_messages.py)
- [test_tool_messages_async.py](https://github.com/langroid/langroid/blob/main/tests/main/test_tool_messages_async.py)
- example scripts: [chat-tree-structured.py](https://github.com/langroid/langroid/blob/main/examples/basic/chat-tree-structured.py),
[chat-tree-structured-simple.py](https://github.com/langroid/langroid/blob/main/examples/basic/chat-tree-structured-simple.py)

fix: Catch `XMLException` not just Pydantic `ValidationError` on mis-generated tools

feat(minor): Simplify `SQLChatAgent` - use helper agent directly instead of via task

fix: rename `llm_response` and `llm_response_async` arg to `message` in all Agent-derived classes.