fixes for events in rf hist_random

[Alexandr-Solovev]

Summary

Fixes a host/device memory ordering hazard in the GPU histogram-based Random Forest training kernel. The two helpers that produce per-node random data for split selection (gen_feature_list and gen_random_thresholds) were filling shared-USM buffers on the host and returning an empty sycl::event, then the caller blocked the host on event.wait_and_throw() before launching the consumer kernel. With shared USM, host writes that race a device kernel on the same allocation can be observed inconsistently — and the empty event provided no actual ordering guarantee against subsequent device work.

What changed

File: cpp/oneapi/dal/algo/decision_forest/backend/gpu/train_kernel_hist_impl_dpc.cpp

gen_feature_list

• Allocate the staging buffer as a plain host ndarray (selected_features_host) instead of shared USM.
• Allocate a device-only buffer (selected_features_com, alloc::device) and copy host → device with assign_from_host, returning the resulting sycl::event (was previously an empty event).

gen_random_thresholds

• Same pattern: host ndarray for the RNG fill, then a device buffer + assign_from_host whose event is returned.

Completeness and readability

• I have commented my code, particularly in hard-to-understand areas.
• I have updated the documentation to reflect the changes or created a separate PR with updates and provided its number in the description, if necessary.
• Git commit message contains an appropriate signed-off-by string (see CONTRIBUTING.md for details).
• I have resolved any merge conflicts that might occur with the base branch.

Testing

• I have run it locally and tested the changes extensively.
• All CI jobs are green or I have provided justification why they aren't.
• I have extended testing suite if new functionality was introduced in this PR.

Performance

• I have measured performance for affected algorithms using scikit-learn_bench and provided at least a summary table with measured data, if performance change is expected.
• I have provided justification why performance and/or quality metrics have changed or why changes are not expected.
• I have extended the benchmarking suite and provided a corresponding scikit-learn_bench PR if new measurable functionality was introduced in this PR.

[Alexandr-Solovev]

/intelci: run

[Copilot]

Pull request overview

Fixes host/device ordering for per-node random data generation in the GPU histogram-based Random Forest training path by ensuring the generated feature lists and thresholds are properly transferred to device memory and ordered via SYCL events.

Changes:

• gen_feature_list: generate selected features on host, then copy into a device allocation and return the copy event.
• gen_random_thresholds: generate random thresholds on host, then copy into a device allocation and return the copy event.
• Training loop: stop waiting on empty events; instead, pass the returned copy events as dependencies into compute_best_split.