Langroid

feat: Recognize LLM-generated Tool/Fn-call when `request` field was forgotten but it is still uniquely recognizable via arguments specific to the Tool.

Fix: gpt4o on Azure, Momento, tool message parsing (525), qdrant docker mode

* gpt4o on Azure: handle their odd model version specification, see `handle_gpt4_model` fn in `azure_openai.py`
* chore: freeze momento vector-db version since new version breaks
* fix: `openai_gpt.py`: use `parse_imperfect_json()` for function arg parsing, to handle various edge cases.
* add docker mode for qdrant, see `QdrantDBConfig.docker`

Feat: Add `FastEmbed` embedding model, see https://qdrant.github.io/fastembed/qdrant/Usage_With_Qdrant/

- Introduces class `FastEmbedEmbeddings` and corresponding config class `FastEmbedEmbeddingsConfig`
- Example usage here: https://github.com/langroid/langroid/blob/main/tests/extras/test_fastembed_embeddings.py

feat: SQLChatAgent allow addressing; update examples

Patch: un-pin `sentence-transformers`, `torch` versions in `pyproject.toml`, to avoid issues with Python 3.12

Bug-fix: ChatAgent.llm_response_async - was giving an error when msg is tool. The code wasn't updated when llm_response() was updated.
Right behavior is to return None when the input is a valid tool msg.