[WIP, do NOT merge yet] ENH: Try fallback to sklearnex-CPU from GPU calls before fallback to sklearn-CPU + document array API behavior

[david-cortes-intel]

Description

Currently, when an operation is attempted on GPU and the operation is not supported by oneDAL, sklearnex will fall back straight away to sklearn on CPU, even if the operation might be supported by oneDAL on CPU.

This PR modifies the logic such that it will first check if the operation on CPU would be supported by oneDAL and fall back to it instead of to sklearn.

Along the way, it updates the docs about array API to reflect the current situation in unambiguous terms, and removes lots of outdated things in that page that no longer apply.

A few notes:

• I'm not sure if oneDAL-CPU should have higher or lower priority than sklearn-arrayAPI. Currently, this will pass non-USM array-API inputs to sklearn without attempting a fallback to oneDAL-CPU, which I think is the most reasonable route.
• I am not entirely sure that the mechanism here won't break any kind of contrived setup that's currently working. Please give it a thorough review.

---

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.

[david-cortes-intel]

/intelci: run

[david-cortes-intel]

/intelci: run

[codecov]

Codecov Report

❌ Patch coverage is 75.00000% with 3 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
sklearnex/_device_offload.py	40.00%	2 Missing and 1 partial ⚠️

Flag	Coverage Δ
azure	80.73% <75.00%> (?)
github	73.23% <75.00%> (+<0.01%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
onedal/utils/_sycl_queue_manager.py	82.50% <100.00%> (+1.67%)	⬆️
sklearnex/_device_offload.py	80.72% <40.00%> (-2.62%)	⬇️

... and 28 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[icfaust]

@david-cortes-intel Thanks for the work! If you wouldn't mind, could you split the docs changes off into a separate PR to simplify review/ focus theme/ speed integration? I know its still in draft, but just looking at the timeline for next week.

[david-cortes-intel]

@david-cortes-intel Thanks for the work! If you wouldn't mind, could you split the docs changes off into a separate PR to simplify review/ focus theme/ speed integration? I know its still in draft, but just looking at the timeline for next week.

Sure, but this is not an urgent thing and not a high priority for the next release. It'd make more sense to first address the issue that we have with allow_fallback_to_host sometimes not working, otherwise the docs would have to describe a situation that's not very logical.

[david-cortes-intel]

Closing in favor of #2707 for the docs part, and leaving the change in config for the future after potential bugs are addressed.