[FEA] Standardize KMeans Instantiations

[tarang-jain]

Simplify and standardize the kmeans public API types:

• Extents: Remove all int (int32) extent overloads; only int64_t remains. This halves the number of template instantiations
• Labels: All predict and fit_predict outputs use uint32_t labels (cluster IDs always fit in 32 bits since n_clusters is int). Removes redundant int and int64_t label overloads from balanced kmeans.
• n_iter: Changed from host_scalar_view<IndexT> to host_scalar_view<int> throughout — iteration count is always small.

These changes are made in the detail templates themselves, so the .cu wrappers are clean pass-throughs with no temporary buffers or type conversions.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[cjnolet]

@tarang-jain these changes will need to be verified through cuml and cugraph. Does Faiss need to be checked also?

[tarang-jain]

Faiss wont be affected.
cuML will be affected yes.