fix(llm-observability): flatten langchain's additional_kwargs

[skoob13]

Problem

OpenAI wrappers send data as:

{
    "role": "assistant",
    "content": "",
    "tool_calls": [
        {
            "type": "function",
            "id": "123",
            "function": {
                "name": "test",
                "args": "{\"a\": 1}",
            },
        }
    ]
}

But the Langchain integration adds the tool calls into the additional_kwargs field:

{
    "role": "assistant",
    "content": "",
    "additional_kwargs": {
        "tool_calls": [
            {
                "type": "function",
                "id": "123",
                "function": {
                    "name": "test",
                    "args": "{\"a\": 1}"
                }
            }
        ]
    }
}

Changes

Flatten additional_kwargs to mirror the OpenAI wrapper to remove the discrepancies.

[Twixes]

Ah, this makes sense – will need to update PostHog/posthog#27626 too