Adjust GPU graph building params

[mayya-sharipova]

cuvs 2025.12 will provide an API for converting HNSW CPU Params to Cagra params.

But for current ES that uses 2025.10 version, we need to adjust params
ourselves. This PR adjust params based on the code from the cuvs library.

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[elasticsearchmachine]

Hi @mayya-sharipova, I've created a changelog YAML for you.

[mayya-sharipova]

New cagra params for GPU index building based on this PR
params.graph_degree = 2 + M * 2 / 3;
params.intermediate_graph_degree = M + M * ef_construction / 256
KnnIndexTester: 8 indexing threads ; Nvidia GeForce RTX 4060, 8Gb memory

---

With this adjustments, we can make GPU indexing/merging up to **3x faster ** than before

gist: 1_000_000 docs; 960 dims; euclidean metric

index_type	index_time(ms)	force_merge_time (ms)	QPS multiple segs	recall multiple segs	QPS single seg	recall single seg
cpu	200730	130129	140	0.94	421	0.91
gpu old	58743	65363	155	0.98	358	0.97
gpu PR	22119	20643	177	0.96	467	0.92

cohere-wikipedia_v2: 934_024 docs; 768 dims; cosine metric

index_type	index_time(ms)	force_merge_time (ms)	QPS multiple segs	recall multiple segs	QPS single seg	recall single seg
cpu	101200	146709	171	0.94	471	0.80
gpu old	44364	48228	172	0.99	367	1
gpu  PR	16298	15961	245	0.99	532	0.98

quora-E5-small: 522931 docs; 384 dims; cosine metric

index_type	index_time(ms)	force_merge_time (ms)	QPS multiple segs	recall multiple segs	QPS single seg	recall single seg
cpu	41025	66514	301	0.98	792	0.93
gpu old	21338	22459	238	0.98	546	1
gpu PR	9270	6532	323	1	769	1

opeanai: 2.6M docs; 1536 dims; dot_product (cosine for GPU)

index_type	index_time(ms)	force_merge_time (ms)	QPS multiple segs	recall multiple segs	QPS single seg	recall single seg
cpu	702350	149302	99	0.85	148	0.84
gpu old	283954	46015	123	0.99	160	0.96
gpu PR	94886	15343	194	0.97	224	0.97

[ChrisHegarty]

Awesome! LGTM

[elasticsearchmachine]

💚 Backport successful

Status	Branch	Result
✅	9.2