Investigate image support with Transformers

[anakin87]

We would like to use some models via Transformers that support multimodal user messages.

• We want to support (some) Image-Text-to-Text models
• The current component (HuggingFaceLocalChatGenerator) might not easy/practical to extend, and it might make sense to develop a dedicated component
• I would not give this investigation high-priority: for multimodal open models, it's better to first focus on Ollama that provides more standardization and does not require GPU; I would also expect users who have GPU to run vLLM (currently in Haystack it can be done via OpenAI with some limitations)