Improve the handling of inputs for TransformersMultiModal

[RobinPicard]

The current approach in which we request users to include the asset tags in their text prompts causes problems for models for which the call to format_chat_template is expected to add them (for instance Qwen2.5VL). In general, we should review the way we turn user inputs into model inputs for this model to make sure it works with different types of models (so we need look at the format_input method of the type adapter).