Support fitted IsolationForest conversion to scikit-learn

[csadorf]

Problem

cuml.accel synchronizes fitted GPU state to the CPU estimator before returning public fitted attributes, falling back to CPU after a GPU fit, or pickling the proxy.

cuml.ensemble.IsolationForest._attrs_to_cpu currently raises UnsupportedOnCPU for every fitted model, leaving the CPU estimator unfitted. Consequently, #8477 cannot fully implement the proxy contract required by #8468 without error-only behavior.

Acceptance criteria

• as_sklearn() succeeds for a fitted cuML IsolationForest.
• Converted score_samples() results agree with cuML within an appropriate floating-point tolerance.
• Converted predictions agree across default, max_features, contamination, and bootstrap configurations.
• Required fitted attributes and scikit-learn fit caches are populated.
• The converted estimator preserves behavior across a pickle round trip.
• cuml.accel can synchronize fitted state for attribute access, post-fit CPU fallback, and pickling.
• Ambiguous sample-count reconstruction fails with a clear error.

The reverse scikit-learn → cuML conversion and populating Treelite data_count are follow-up work.

[JulienAu]

Thanks for tracking this. I'm on it, and I'll get the PR up in the next few days, covering all the acceptance criteria above including the fail-loud check on ambiguous sample-count reconstruction. Could you assign this to me?

[csadorf]

Thank you!