Add support for WARP

[pau-mensa]

Summary

• Add a new WARP index backend wrapping xtr-warp-rs and export it from pylate.indexes
• Implement add_documents, remove_documents (with auto-compaction), and search with optional document ID filtering (shared or per-query)
• Add update_documents and compact as new public methods not present in the base Index class. update_documents replaces embeddings in-place without delete+add, compact physically removes tombstoned passages, empty centroids and reclaims disk space
• Expose all xtr-warp-rs hyperparameters via two config dataclasses (WARPSearchConfig, WARPIndexingConfig) to keep the constructor clean
• Extract shared embedding conversion utilities (convert_embeddings_to_torch) into pylate/indexes/utils.py for reuse across WARP and FastPLAID
• Add xtr-warp-rs as an optional dependency
• Add 12 focused WARP unit tests

Tests

• python -m pytest tests/test_warp.py

[pau-mensa]

Notes

• WARP supports metadata filtering now, but I did not add it because I saw that the other indices don't have it
• WARP also supports streamed indexing, which can cut VRAM usage by 66%, but that would require changing things that seem a bit out of scope. Still, it is a powerful feature, let me know if you want me to add it