Modernize cuda.ml for CUDA 13.3 and RAPIDS cuML 26.6

[t-kalinowski]

Summary

Modernizes cuda.ml for CUDA 13.3 and RAPIDS cuML 26.6 while retaining version-guarded support for the existing cuML 21.x build paths.

The installation flow can now bootstrap RAPIDS cuML from pip wheels when a compatible GPU and CUDA toolkit are available. CRAN and non-GPU environments continue to receive a stub-only build with clearer diagnostics.

Why

The native integration assumed CUDA 11, cuML 21 APIs, the older Treelite/FIL interface, and a manually provisioned RAPIDS prefix. Current RAPIDS releases changed API signatures, resource handling, packaging layouts, and several model structures.

These differences caused compilation failures, incorrect host/device transfers, ignored normalization, invalid FIL probability output, stale incremental builds, and GPU CI packages being configured without GPU access.

Public-facing changes

• Adds automatic RAPIDS cuML bootstrap through uv or Python/pip for CUDA 12 and 13.
• Improves installation diagnostics and documents CUDA_HOME, CUML_PREFIX, CUML_BOOTSTRAP, and related configuration options.
• Preserves normalize_input = TRUE behavior for lasso, elastic net, and ridge models on modern cuML.
• Corrects FIL probability output for multiclass models using max_index postprocessing.
• Correctly forwards KNN n_bits and updates behavior for modern KNN implementations.
• Stabilizes TSVD component signs across cuML versions.
• Produces clear errors for capabilities no longer supported by modern cuML:
  • non-default legacy FIL loading controls;
  • IVFSQ KNN;
  • random forest serialization without Treelite/FIL;
  • random projection when its cuML C API is unavailable.
• Updates installation and package documentation for the current GPU and RAPIDS requirements.

Internal changes

• Adapts clustering, decomposition, linear model, random forest, SVM, KNN, t-SNE, and UMAP bindings to modern cuML and RAFT APIs.
• Retains version guards for the bundled cuML 21.x paths.
• Replaces ambiguous Thrust copy dispatch with explicit host/device-aware asynchronous CUDA copies.
• Migrates FIL and Treelite integration to the current forest model, import, prediction, and serialization APIs.
• Uses modern RAFT stream resources and RAPIDS logging types.
• Adds out-of-tree CMake builds, reliable incremental rebuilds, RAPIDS library discovery, and runtime library paths.
• Moves GPU-linked package installation into the GPU-visible CI job.
• Updates the CI image to CUDA 13.3 on Ubuntu 24.04 and verifies cuML linkage before running GPU tests.
• Regenerates documentation with roxygen2 8.0.

Validation

• CRAN check on R release
• CRAN check on R devel
• Docker CI image build
• GPU-linked installation and full GPU test suite
• Targeted FIL probability tests with RAPIDS 26.4 and Treelite 4.7
• Regression tests for normalization, FIL controls, KNN compatibility, serialization, and stub-only behavior

All GitHub Actions checks pass.

[dfalbel]

It's important to keep installing the package in the docker image to avoid using the expensive GPU runner for compiling the C++ source.