Langroid

* DocChatAgentConfig: new `filter` param to be used to constrain searched document-set.
This gets used inside `get_semantic_search_results` and is passed to `self.vecdb.similar_texts_with_scores()`

* LanceDocChatAgent: version of DocChatAgent that leverages LanceDB filtered vector-search and full-text search
(see [`tests/main/test_lance_doc_chat_agent.py`](https://github.com/langroid/langroid/blob/main/tests/main/test_lance_doc_chat_agent.py))

* Task: new param `max_stalled_steps` (default = 3): If stalled for more than this many steps, quit
* Agent: friendly warning on mis-named tool
* DocChatAgent: make relevance extractor optional

* OpenAI streaming: use retry-with-exp-backoff
* ChatAgent.llm_response_forget: check msg history length
* LanceDB update to latest, which natively supports nested Pydantic models (so no need to use our workaround)
* Document.metadata.id: change type from str|None -> str, with default = ""

Task constructor takes new param `allow_null_result`, which defaults to True,
and this setting is backward-compatible.
When True, it means an empty or DO-NOT-KNOW result is allowed as a step or task-level result.
For some scenarios it is useful to set it to False; this retains the current pending message
so it can be passed on to a sub-task, or returned as the result of a task.

Agent.base: agent_response(): improve err msg when missing/wrong field in ToolMessage
(Previously only the top-level tool name was being returned; this improves it to show the actual missing/wrong field)