[feat] Integrate NanoBeIR datasets

[ArthurCamara]

As discussed in #2848 (comment), This PR adds a new Evaluator based on the NanoBEIR collection of datasets.

It creates one InformationRetrievalEvaluator for each dataset, and aggregates the results accordingly.

Example:

from sentence_transformers import SentenceTransformer
from sentence_transformers.evaluation import NanoBEIREvaluator

# Load a model
model = SentenceTransformer('all-mpnet-base-v2')

datasets = ["QuoraRetrieval", "MSMARCO"]
query_prompts = {
"QuoraRetrieval": "Instruct: Given a question, retrieve questions that are semantically equivalent to the given question\nQuery: ",
"MSMARCO": "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery: "
}

evaluator = NanoBEIREvaluator(
dataset_names=datasets,
name="NanoBEIR",
query_prompts=query_prompts,
)

results = evaluator(model)
'''
NanoBEeIR Evaluation of the model on ['QuoraRetrieval', 'MSMARCO'] dataset:
Evaluating NanoBeIRNanoQuoraRetrieval
Evaluating NanoBeIRNanoMSMARCO

Average Queries: 50.0
Average Corpus: 5044.5

Aggregated for Score Function: cosine
Accuracy@1: 39.00%
Accuracy@3: 57.00%
Accuracy@5: 66.00%
Accuracy@10: 77.00%
Precision@1: 39.00%
Recall@1: 34.03%
Precision@3: 20.67%
Recall@3: 54.07%
Precision@5: 15.00%
Recall@5: 64.27%
Precision@10: 8.90%
Recall@10: 75.97%
MRR@10: 0.5004
NDCG@10: 0.5513
Aggregated for Score Function: dot
Accuracy@1: 39.00%
Accuracy@3: 57.00%
Accuracy@5: 66.00%
Accuracy@10: 77.00%
Precision@1: 39.00%
Recall@1: 34.03%
Precision@3: 20.67%
Recall@3: 54.07%
Precision@5: 15.00%
Recall@5: 64.27%
Precision@10: 8.90%
Recall@10: 75.97%
MRR@10: 0.5004
NDCG@10: 0.5513
'''
logger.info(evaluator.primary_metric)
# => "cosine_ndcg@10"
logger.info(results["mean"][evaluator.primary_metric])
# => 0.5512516989358924

(Note that this depends on #2951)

[tomaarsen]

Although the Be portion obviously stands for Benchmark, I think the abbreviated "BEIR" is usually fully capitalized, so I'd like to propagate that in this PR as well.

[tomaarsen]

I'm experimenting with having all outputs in the final dict, rather than a nested dict. This way, people can use any value from the evaluator to guide their e.g. early stopping. It should also match the SequentialEvaluator performance, even though the results from the NanoBEIR are now a bit hectic (i.e., one massive dict).

I hope it's okay if I push into this PR!