fix: deactivate function calls when we already have the result to avoir multiple round trips

[nlenepveu]

Issue:

When using output_model in combination with a tool, the LLM repeatedly requests Function Calls, leading to multiple round trips. Here’s a snippet that reproduces the behavior: https://gist.github.com/nlenepveu/dceb6e0f27cf1d897f3768736070f1aa

Fix:

While the tool_call_limit parameter limits the total number of tool executions, there are instances where the LLM requests for multiple tool calls within a single response, and this parameter cannot enforce a single round trip. Since the LLM’s behavior is unpredictable in terms of how many calls it requests in a single response, the current tool_call_limit parameter is insufficient.

The proposed fix is to modify the default behavior to automatically set the tool_choice to "none" after the first round trip. This ensures the LLM doesn't attempt additional tool calls once it has already received the necessary results from the initial tool execution.

[nlenepveu]

I'm starting to doubt whether this is the right way to change the behavior identified in my example. We must still allow the LLM to make function calls after the first round trip, because then it might need a different tool.. Let me know your thoughts on this issue.