[FEA] Add `cuml.accel` support for IsolationForest

[btepera]

Is your feature request related to a problem? Please describe.
cuML now has a GPU-accelerated Isolation Forest estimator (cuml.ensemble.IsolationForest, added in #8226), but cuml.accel does not dispatch to it. Anomaly detection users with large tabular data would benefit from IsolationForest being added under the cuml.accel umbrella of estimators.

Describe the solution you'd like
Create a cuml.accel proxy for sklearn.ensemble.IsolationForest backed by cuml.ensemble.IsolationForest, following the same pattern used for other estimator proxies.

When complete, I should be able to run fit() and predict() on GPU:

%load_ext cuml.accel
from sklearn.ensemble import IsolationForest

clf = IsolationForest(n_estimators=100, random_state=0)
clf.fit(X)

After adding to cuml.accel, we should also update the Accelerated Estimator Support section of our docs.

[adityaanikam]

Confirmed cuml.ensemble.IsolationForest is there and the ProxyBase pattern in accel/_overrides/sklearn/ensemble.py, the same one RandomForestRegressor and RandomForestClassifier use, should drop in cleanly for this. One thing worth flagging since it might affect where you want this tracked: the repo has moved from rapidsai/cuml to NVIDIA/cuml. I'd like to put up the proxy.

[JulienAu]

I'd like to work on this one. Plan: add the IsolationForest proxy in python/cuml/cuml/accel/_overrides/sklearn/ensemble.py following the existing estimator-proxy pattern (and using #8472's HDBSCAN addition as the reference for the current conventions), wire it into the accel estimator registry, add the accel tests for fit/predict/decision_function/score_samples parity against sklearn.ensemble.IsolationForest, and update the Accelerated Estimator Support docs section.

One question before I start, since #8420 tracks deferred interop work: if some sklearn surface of IsolationForest is not yet backed by cuml.ensemble.IsolationForest, is the preferred behavior for the proxy to fall back to CPU for those calls (as other proxies do), or should I defer those to the follow-up issue?

[JulienAu]

@adityaanikam you commented first and I missed it when I claimed above, sorry about that. The proxy is yours; I'm happy to review when you put it up.

One technical heads-up from reading the code that may save you a debugging session: IsolationForest._attrs_to_cpu currently raises UnsupportedOnCPU (fitted-model conversion is deferred work tracked in #8420), and ProxyBase.__getattr__/__reduce__ call _sync_attrs_to_cpu unconditionally. So after a GPU fit, accessing any *_ attribute (offset_, max_samples_, estimators_) or pickling the proxy will surface that raise unless the proxy handles it, for example with _not_implemented_attributes plus a guarded sync. Note also that with no attributes synced, is_fitted(self._cpu) stays false, which affects the nicer-error path and check_is_fitted.

To keep the two pieces complementary rather than competing: I'll take the interoperability items in #8420 (fitted-model conversion or its documentation, and the NotFittedError item), which is exactly the machinery your proxy would then get for free. I'll comment there.