Replace iris examples and test fixtures

[t-kalinowski]

Summary

• Replace deprecated iris examples across the public documentation with compact datasets from modeldata, and use Old Faithful for the release post's concise k-means demonstration.
• Replace the internal iris-based GPU test fixture with Palmer penguins, preserving coverage for multiclass modeling, clustering, serialization, and dimensionality reduction.
• Update the tidymodels examples to use current native-pipe APIs, add the optional dataset dependencies, and regenerate the README, function reference, vignettes, and pkgdown site.

Public-facing changes

All user-facing examples now avoid iris and use complete datasets without missing-value cleanup. The release post keeps installation focused on cuda.ml and assumes optional example packages are available. Other documentation uses ordinary numeric literals and set.seed() where reproducibility matters.

Internal changes

GPU tests use a shared, scaled Palmer penguins fixture. Dataset-dependent assertions now account for component-sign ambiguity and compare clustering quality using inertia and adjusted Rand index. There are no public API changes.

Testing

• Rscript -e 'devtools::test(reporter = "progress")' (424 passed, 9 skipped)
• R CMD build /home/tomasz/github/mlverse/cuda.ml
• _R_CHECK_CRAN_INCOMING_=false _R_CHECK_SYSTEM_CLOCK_=false R CMD check --no-manual --as-cran cuda.ml_0.4.0.tar.gz (Status: OK)
• Rscript -e 'rmarkdown::render("README.Rmd")'
• Rscript -e 'pkgdown::build_site()'
• Rendered the Old Faithful k-means plot with Quarto against the GPU backend.