Automatically remove the string "<|image_sentinel|>" from chat outputs

[rhennigan]

This occasionally appears in chat outputs when using gpt-4-vision. It's clearly a special token used by OpenAI, since the vision model can't read it like other models can:

In[1]:= AssociationMap[
    Function[
        LLMSynthesize[
            "Repeat the following string back to me:\nMy favorite token is <|image_sentinel|>!",
            LLMEvaluator -> <| "Model" -> #1, "MaxTokens" -> 64 |>
        ]
    ],
    { "gpt-3.5-turbo", "gpt-4", "gpt-4-turbo-preview", "gpt-4-vision-preview" }
]

Out[1]= <|
    "gpt-3.5-turbo"        -> "My favorite token is <|image_sentinel|>!",
    "gpt-4"                -> "My favorite token is <|image_sentinel|>!",
    "gpt-4-turbo-preview"  -> "My favorite token is <|image_sentinel|>!",
    "gpt-4-vision-preview" -> "\"My favorite token is !\""
|>

The fact that it occasionally appears in chat outputs is almost certainly a bug on OpenAI's end. However, it's simple enough to just delete it from chat outputs as a workaround.