Multi-GPU out of core KMeans progress tracking issue

This issue aims at tracking progress on the multi-GPU out of core KMeans implementation and benchmarking.

Please tell me if anything was forgotten. Also, if you add a new issue/PR, please list it here and link this issue as its parent so we can keep track of everything that is going on.

Main Tasks

• Multi-GPU batched KMeans core (#2017)
• Multi-GPU KMeans implementations merge + cleanup (#2015)
• SNMG Batched KMeans Python API (#2154)
• Handle multiple partitions per rank/worker + API merges (#2066)
• Handle multiple partitions per rank/worker (cuML side) (rapidsai/cuml#8201 and rapidsai/cuml#8084)
• Evaluate KMeansPlusPlus Init For Out of Core (#2224)
• SNMG Batched KMeans Python API benchmark (#2149)
• Large scale multi-node Dask KMeans benchmark (rapidsai/cuml#8198)

Improvements and bugfixes

• Out-of-core K-means improvements (#2292)
• Apply ABI Breaking Changes for KMeans (#2147)
• Reuse Precomputed Norms for Inertia Computation (#2057)
• Evaluate Prefetch for Batched Kmeans (#1917)
• Clamp max_iter in OOC KMeans test (#2256)
• KMeans ignores sample_weight when computing inertia and score (rapidsai/cuml#8530)
• KMeans.transform returns squared distances instead of Euclidean distances (rapidsai/cuml#8536)

Review the unchecked tasks in this tracking issue, starting with the linked issues #2224, #2149, #8198, and #2292. Use the listed cuML issues and related benchmarks to understand the remaining multi-GPU out-of-core KMeans work. Done means completing a scoped task and updating this checklist with the relevant issue or pull request link.