Simplify retrieval

[Muennighoff]

• Remove beir dependency - This makes installation easier due to less dependencies
• Remove outdated Retrieval evaluator - Afaik this evaluator was never used so it's confusing duplicate code
• Remove beir multinode code - As outlined here this can be much easier done on the user side and the current implementation is too hacky
• Add & adapt some classes from beir, mainly the hf dataset loader & the evaluator
• Add all beir retrieval datasets to mteb on the hub and load them from there - This ensures consistency and allows people to inspect datasets much easier (previously they were loaded not from hf)

I tested with komninos that all English Retrieval results for the datasets changed remain the same. It would be great if all of you could take a look at the PR, as this is the biggest change to MTEB since release.

I also tested to make sure the BeIR-PL tasks & Korean Retrieval tasks still work (cc @kwojtasi @taeminlee)

[imenelydiaker]

Globally, everything is good. I just left a comment about a test and a function call.
I've also added some datasets revision IDs, I guess you'll change them if you update the datasets.

[Muennighoff]

@NouamaneTazi @loicmagne @tomaarsen Any thoughts? 😊

[loicmagne]

Looks good, great work!