feat: MultiModalRetriever

[ZanSara]

Related Issue(s):

• Closes Verify compatibility between Data2VecVision models and existing retrievers #2865
• Closes Load CLIP Models into a Retriever for image retrieval #2857
• Related to Add support for images #2418

Proposed changes:

• Create a multi modal retriever by generalizing the concepts introduced by TableTextRetriever
• It introduces a stack of new subclasses to support such retriever, such as MultiModalEmbedder)
• Note that this Retriever will NOT be tested for working in pipelines, but only to work in isolation. It will also, most likely, stay undocumented. See Add support for images #2418 for the rationale.

Additional context:

• As mentioned in the original issue, an attempt to generalize TableTextRetriever quickly proved too complex for the scope of this PR.
• Rather than modifying an existing Retriever with the risk of breaking working code, I opted for cloning the class and its stack of supporting classes and perform the changes needed to support N models rather than just 3.
• A later goal is to be able to perform table retrieval with MultiModalRetriever and use its stack to dispose of TriAdaptiveModel, BiAdaptiveModel and maybe AdaptiveModel itself, along with their respective helpers (custom predictive heads, custom processors, etc).

Additional changes:

• Soon I realized that with image support we need to generalize the concept of tokenizer. So I renamed haystack/modeling/models/tokenization.py -> haystack/modeling/models/feature_extraction.py, created a class called FeatureExtractor and used it as a uniform interface over AutoTokenizer and AutoFeatureExtractor

Pre-flight checklist

• I have read the contributors guidelines
• If this is a code change, I added tests or updated existing ones
• If this is a code change, I updated the docstrings

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

Changes addressed