DimReducer: add prefix param + tests (v0.3.4)

[jrosenfeld13]

Summary

• Add a prefix param to DimReducer so output columns can be named {prefix}_{i} instead of the hardcoded dim_{i}. Default is "dim", so existing behavior is unchanged (backward compatible).
• Cast reducer input to float32 (matches t-SNE/UMAP internals; enables cuML/GPU backends). Note: PCA on float64 inputs will produce slightly different numeric results than 0.3.3.
• Add DimReducer test coverage (previously zero): default/custom prefix, get_feature_names_out parity, invalid-method validation, pandas backend, and pca/tsne/umap methods.
• Trim docstrings, add a cuML drop-in note, and make the docstring examples runnable (doctest passes).
• Bump version 0.3.3 -> 0.3.4.

Notes

• Method validation was kept eager in __init__ (restored + tidied into a _VALID_METHODS constant), so this release has no API-observable behavior change beyond the additive prefix param and the float32 cast.
• Pre-existing (out of scope): base-class mutates feature_names in fit, which trips sklearn clone/check_estimator. Tracked separately.

Test plan

• uv run pytest tests/test_feature_transformers.py — 11 passed, 6 skipped (optional deps)
• DimReducer doctest — 7 attempted, 0 failed
• Confirm CI green on the PR

Made with Cursor