Rewrite kernel profiler for correct tree nesting and GPU timing

[Alexandr-Solovev]

Rebuild the profiler tree from per-thread stacks tracked in the singleton instead of a global level counter. Regular tasks now record a parent index at push time (the caller thread's innermost open task, with fallback to the innermost regular task across threads for worker-thread threading tasks); level derives from parent + 1. This fixes the case where a task opened on one thread and closed on another (or threading tasks under a regular parent) produced garbled nesting.

Merge phase now keys sibling collapse on (parent_idx, name) via an unordered_map, so it is O(n) and correctly matches only true siblings. Levels-and-names on their own were ambiguous: two tasks at the same depth under different parents could collapse into one.

Cache ONEDAL_VERBOSE once per process in a function-local static instead of a namespace-scope volatile int. The previous form gave every TU that included the header its own copy, so daal_verbose_val observed different values across .so boundaries.

Move the once-per-process suppression sets for the threading logger and tracer messages onto the profiler singleton so their lifetimes are bounded to the profiler, not the process; hoist unique-name checks under the existing mutex.

On the oneAPI side, replace the start_task(name, queue) pattern with a queue_sync_guard RAII wrapper. The guard issues wait_and_throw() before start_task() and again on scope exit so recorded durations reflect device-side execution rather than host enqueue. The pre-region wait happens before the timer, so already-in-flight work on the queue is not charged to the region. When the profiler is disabled the guard is a nullptr no-op and no waits are issued.

Uniquify the local variable name via LINE so multiple profiled scopes in the same block coexist; drop the ternary-of-mixed-types pattern and gate each START_TASK* macro on the enable predicate directly. SERVICE task variants gate their logger side-effects on is_service_debug_enabled (ONEDAL_VERBOSE=5) rather than is_logger_enabled (=1/4), matching the pre-refactor behavior — otherwise hot paths like table2ndarray would flood the output under LOGGER mode.

Validated: make onedal_dpc -j112 clean; CPU + GPU cov_dense_batch and kmeans_lloyd_dense_batch under ONEDAL_VERBOSE=3 produce well-formed analyzer trees with correct parent-child nesting and durations.

Description

---

Checklist:

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.

[david-cortes-intel]

It could explain what a TU is.

[david-cortes-intel]

Are macros from this header used somewhere?

[david-cortes-intel]

Throwing on a destructor is undefined behavior. It should be catched regardless of this macro.