Neural-cherche

Update of the ColBERT Ranker:
- Avoid raising an error when there are duplicates documents with two distinct ids.

This release reduce BM25 memory usage by casting sparse matrice values to float32. Avoid copy of the sparse matrix when computing BM25 scores.

Update dependancies, make BM25 retriever normalisation compatible with previous version of sklearn.

Update LeNLP dependancy version in order to run without error on older Ubuntu version.

The version 1.3.0 introduce:

- A new BM25 retriever powered by LeNLP vectorizer written in rust. SOTA.
- An updated TfIdf retriever with LeNLP TfidfVectorizer by default which is written in rust, not SOTA but really fast.

- Breaking change with the evaluation code. The evaluation module is much simpler and will handle duplicates queries.
- Update of the documentation.

- ColBERT Retriever is now available (complementary to the ColBERT ranker).
- Improved default settings for every models.
- Attention mask added to models.
- ColBERT and SparseEmbed pre-trained checkpoints on HuggingFace: raphaelsty/neural-cherche-colbert and raphaelsty/neural-cherche-sparse-embed.
- Improved ranking loss.
- Addition of benchmarks.

Overall, this version makes easier to fine-tune ColBERT, SparseEmbed and Splade and achieve excellent results without default parameters.