[FEA] Document HNSW-to-CAGRA migration for Faiss Python and cuVS Bench #2879

Problem

Faiss Python users migrating from HNSW to CAGRA must reconstruct the cuVS C++ HNSW heuristics manually. The CAGRA cuVS guide shows from_hnsw_params in a C++ conversion example, but does not explain the effective settings or how to apply them in Faiss Python and cuVS Bench. HNSW M is not directly equivalent to CAGRA graph_degree.

Requested change

Add a Faiss Python migration example that:

• Takes HNSW M and efConstruction, dataset size and dimension, and distance metric.
• Explains SIMILAR_SEARCH_PERFORMANCE and SAME_GRAPH_FOOTPRINT and shows the resulting graph degree, intermediate graph degree, graph builder, and builder-specific settings.
• Provides runnable Faiss Python and equivalent cuVS Bench YAML, identifying any settings the interfaces cannot represent.
• Separates GPU-built CPU IndexHNSWCagra search from native GPU CAGRA search, including conversion options and the applicable search sweeps.
• Explains that heuristics provide starting builds, requiring measurement of exact recall and latency or throughput at matched recall.

For M=64 and efConstruction=512, the current heuristics produce GD44/IGD192 and GD128/IGD192 respectively. The builder is NN-Descent below one million vectors and IVF-PQ otherwise. Include the full builder configuration, not just degrees, and document applicable versions.

References

• https://docs.nvidia.com/cuvs/user-guide/api-guides/indexing-guide/cagra
• https://github.com/NVIDIA/cuvs/blob/main/cpp/src/neighbors/cagra.cpp
• https://github.com/facebookresearch/faiss/blob/main/faiss/gpu/GpuIndexCagra.h
• https://github.com/NVIDIA/cuvs/blob/main/python/cuvs_bench/cuvs_bench/config/algos/faiss_gpu_cagra.yaml

Related broader workflow: #2730.