llm-tool / VLMEval does not build prompt correctly when used with images/video

[davidkoski]

In LLMTool.swift it doesn't build the prompt correctly when there are images or video:

        let messages: [[String: Any]] =
            if !images.isEmpty || !videos.isEmpty {
                [
                    [
                        "role": "user",
                        "content": [
                            [
                                "type": "text",
                                "text": generate.system,
                            ]
                        ]

It looks like this was inadvertently added in #206 -- it needs to include the use prompt. I think this should probably generate two dictionaries and this first one should be role: system.

I think we can fix this in #257