Return the chat messages as input_text when multimodal pipelines return tensors (#50816)
* Return the chat messages as input_text when multimodal pipelines return tensors
With a chat input and return_tensors=True, ImageTextToTextPipeline and
AnyToAnyPipeline returned the internal Chat wrapper as input_text instead
of the list of messages, as the text return types already do.
* Limit the any-to-any regression test to two thinker tokens