[mieb] Any2TextMultipleChoice Abstask&Evaluator & four tasks in CV-bench

[gowitheflow-1998]

Adding:

1. Any2Text Multiple Choice Abstask and Evaluator
2. four tasks from CV-bench that uses above (count, relation, depth, distance).

the logic is first grabbing unique candidate choices from the task (e.g., the 788 questions in the count task only has 17 unique candidates), encode them, store in a {candidate: embedding} dict. In scoring time, we query the dict and do similarity matching. This gets rid of the need to repetitive encoding, and makes it easier for queries with different number of candidates that can't be supported by AbsTaskImageTextPairClassification atm.

Checklist

• Run tests locally to make sure nothing is broken using make test.
• Run the formatter to format the code using make lint.

[gowitheflow-1998]

also, this and Any2AnyRetrieval can be extended to audio and video for moeb at some point I think!

[isaac-chung]

Great work with the AbsTasks! Mostly small non-blocking comments.