Fix prefetch-dataset hang on GPU

[ethanglaser]

Description

Dataset prefetching with --prefetch-datasets was hanging all subsequent GPU benchmarks because multiprocessing.Pool defaults to fork, which corrupts the SYCL/Level Zero GPU runtime when child processes exit. Switched to spawn start method to avoid this.

Still more work needed on this

Via claude

---

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.

[ethanglaser]

GPU run with some progress but still issues: http://intel-ci.intel.com/f132cfc2-18cc-f185-949b-a4bf010d0e2d
Verifying no issues on CPU side: http://intel-ci.intel.com/f132cfa6-f997-f1ab-a141-a4bf010d0e2d

[david-cortes-intel]

@ethanglaser Are you sure this is about sycl runtimes and not openmp? If it was about openmp, it could be fixed by switching whatever backend the python process is picking towards intel's or LLVM's, which can work in forked processed. That should be relatively easy if using conda.