feat: add TextChunk as possible content type for Assistant message

[gcalmettes]

Currently, the AssistantMessage class only accepts str as type. It is the only role only accepting nothing else than str (or None) as content.

This means that payload generated when using the Mistral tokenizer via external systems (e.g. vllm) also accepting TextChunk as content for the Assistant message, will break:

"messages": [
    {
        "role": "system",
        "content": [
            {
                "type": "text", "text": "\nCurrent model: pixtral-12b-2409\nCurrent date: 2024-09-23T07:14:36.737Z\n\nYou are a helpful assistant. You can help me by answering my questions. You can also ask me questions."
                }
        ]
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "Does the animal in the first image could live in what is describe in the second image ?"
            },
            {
                "type": "image_url",
                "image_url": {
                "url": "https://some-urls"
                }
            },
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://some-urls"
                }
            }
        ]
    },
    {
        "role": "assistant",
        "content": [
            {
                "type": "text",
                "text": "Some response from Pixtral."
            }
        ]
    },
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": "What time is it ?"
            }
        ]
    }
]

This MR adds TextChunk as possible content type for the AssistantMessage class.

[gcalmettes]

@jean-malo @patrickvonplaten any chance of this PR to be reviewed on your side ? This still causes some incompatiblity when querying mistral with the openAI client (since on their side textChunks are also valid for assistant messages, like for the other roles)

Thanks !

[gcalmettes]

related: vllm-project/vllm#12859

[juliendenize]

Done in another PR !