Fix single-GPU refine CUDA IPC failures

[jdickerson95]

Summary

• Run single-GPU core_refine_template / core_inspect_template in-process instead of spawning a child process after CUDA init
• Avoids pidfd_getfd: Operation not permitted / empty torch.cat failures seen during optimize_template on one GPU
• Multi-GPU paths still use run_multiprocess_jobs
• Add unit tests covering single- vs multi-device dispatch

Test plan

• pytest tests/backend/test_core_refine_template.py
• Re-run single-GPU optimize_template on the failing machine/config
• Spot-check multi-GPU refine still dispatches via multiprocessing

Made with Cursor