Add similarity search backend to Retrieval tasks

[Samoed]

Description of the feature

As #3192 added possible different backends to the cache wrapper, we can add similar backends to retrieval tasks.

It can have interface like

class SearchBackend:
    def add_document(
        self,
        embeddings: Array,
        idxs: list[str],
    ) -> None:
        """Add documents to the search backend.
        
        Args:
            embeddings: Embeddings of the documents to add.
            idxs: IDs of the documents to add.
        """
        
    def search(
        self,
        embeddings: Array,
    ) -> list[tuple[str, float]]:
        """Search the backend for the given embeddings.

        Args:
            embeddings: Embeddings to search for.

        Returns:
            List of tuples containing document IDs and their relevance scores.
        """

    def save(self, **kwargs):
        ...

    def load(self, **kwargs):
        ...

And then this can be used with SearchEncoderWrapper as:

model = SearchEncoderWrapper(
    mteb.get_model(...),
    backend=SearchBackend(),
)

This method wouldn't require changes for other search models (pylate), but propose change for the encoders

CC @orionw

[orionw]

+1, this would be quite nice and mirror what is going on in pylate and in transformers with their attention variants.

At some point it could wrap the many different ways people use backends, e.g. variants of faiss etc.