1.0: Separating vector store from the Index

[ashvardanian]

This PR brings the first significant breaking change to USearch.

New Serialization Format and External Storage

From now onwards, the serialized dense index files are prepended by a .bbin matrix containing type-punned vectors.
It makes it easy to debug and check the contents of each index and potentially reconstruct it.
Moreover, this allows us to integrate the underlying index_gt in places like UStore, where the vectors would be stored externally in a separate collection.
Those collections are often stored on disk - in external memory. Random reads can be expensive, so we now allow batch-evaluated metrics to let you prefetch all the relevant vectors simultaneously, resulting in major performance improvements.

Minor Tweaks

• Batch-evaluated metrics for external prefetching.
• Allow swapping the metric from Python at runtime.
• Switching to brute-force search on tiny collections.
• Search results across different bindings now have a simpler form.
• Switching from JavaScript to TypeScript interface.
• Improved Swift bindings to support both slices and arrays.
• Fix memory alignment issues with atypical vector sizes and SIMD loads.

[ashvardanian]

🎉 This PR is included in version 0.23.0 🎉

The release is available on GitHub release

Your semantic-release bot 📦🚀