Add support for images

[ZanSara]

Problem
So far Haystack has been focusing strongly on text-only search. However, the same architecture is likely to be effective on other mediums, such as images.

This epic tracks the implementation of support for image indexing and retrieval in Haystack.

Process

• Research on the topic
  • Candidate models for image retrieval identified: CLIP and Data2Vec
  • Data2VecVision has been chosen due to the existence of sibling models for text and audio which would allow us to have comparable embeddings across different document types later on (see ImageRetriever #2445).
• Investigate and eventually adapt how Haystack loads models from HF.
  • Simplify language_modeling.py and tokenization.py #2703
• Create a primitive MultiModalRetriever that works in isolation (NOT in a pipeline) with at least one single document store, and add tests.
  • Verify compatibility between Data2VecVision models and existing retrievers #2865
  • Load CLIP Models into a Retriever for image retrieval #2857
  • feat: MultiModalRetriever #2891
• Make MultiModalRetriever work on all docstores, if possible and not too time consuming.
  • Verify support for image retrieval by all document stores #2866
• Extend primitives (Document, Answer, etc) to account for the new data type, and add tests. (WARNING: Big task! Changes might not be as big as initially assumed)
  • Generalize primitives #2867
• Make the MultiModalRetriever work in a query pipeline, and add tests.
• Make the MultiModalRetriever work in a indexing pipeline, and add tests.
• Create a tutorial and some documentation

Later steps (not in order of importance, not blocking each other):

• Implement training and evaluation on image retrieval task
  • Training and evaluation for MultiModalRetriever #3410
• Make MultiModalRetriever work on the REST API and/or make a separate demo for it (a mixed media one would be super-cool tho)
  • Create a multi-modal demo #3623
• Image-to-text conversion, mainly designed for indexing. Note: this is independent from the rest of the changes, could be even picked up now by some brave external contributor.
  • ImageToText & AnswerToImage #2444