feat: let LLM choose whether to retrieve context

[lsorber]

Changes:

• Add support for tool use to the llama.cpp LiteLLM integration.
• The rag or async_rag functions now update the message history with tool use and the assistant response.
• Let LLM choose whether to retrieve context through tool use. Details:
  • Tool use is skipped if RAG context is already provided.
  • Llama.cpp models cannot stream the response with tool_choice="auto" (which lets the LLM choose whether or not to call a tool), so we add a workaround that forces tool use but lets them skip tool execution with a skip=True argument.
  • To help llama.cpp models plan their tool use response, the final message is extended with the tools' JSON schema.
• Fix registering of llama.cpp model_info with LiteLLM.
• Improve RAG generation test cases.
• Improve LLM-embedder test coverage.