Make GPU/NPU test collection safe on CPU-only hosts

Summary

Some GPU/NPU test modules perform device setup at import time. On CPU-only hosts, full pytest collection can fail before markers such as -m "not gpu" have a chance to deselect those tests.

Why this matters

CPU-only CI and local development should be able to collect and run the CPU test suite without initializing CUDA/NPU runtime state. Import-time device setup also makes failures harder to diagnose because collection fails before the test body runs.

Suggested fix

• Avoid creating DeviceManager() or similar accelerator-dependent objects at module import time in test files.
• Move accelerator setup into GPU-marked test functions or fixtures.
• Skip inside those fixtures when the required accelerator runtime is unavailable.
• Verify that pytest tests/ -m "not gpu" can collect and run on CPU-only hosts.

Expected outcome

CPU-only test runs should be able to execute non-GPU tests without CUDA/NPU initialization side effects during collection.

Start with pytest tests/ -m "not gpu" on a CPU-only host, then inspect test modules under tests/ for DeviceManager or other accelerator-dependent setup performed during import. Move that setup into GPU-marked tests or fixtures, skip when the runtime is unavailable, and confirm non-GPU collection and execution no longer initialize CUDA/NPU state.