Fix MIG-related tests in cuda.core.system

[mdboom]

Fix bugs in new MIG-related functionality in cuda.core.system.

[mdboom]

Just removed this test altogether since it isn't universally true and finding a way to write all of the exceptions seemed a little pointless.

[rwgk]

Do you know if this could also affect "Orin"?

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.

[rwgk]

Cursor GPT-5.4 Extra High Fast

---

Findings

• High cuda_core/cuda/core/system/_system.pyx:105, cuda_core/cuda/core/_device.pyx:1003, cuda_core/tests/system/test_system_system.py:43
  The Thor/MIG count mismatch from issue 309 is still unfixed. system.get_num_devices() still returns NVML's physical-device count whenever NVML is available, and cuda.core.Device.get_all_devices() still uses that value to enumerate CUDA devices. Deleting test_num_devices() removes the direct assertion on that bug, but does not fix the underlying behavior, and test_devices() still goes through the same bad path on a topology where NVML sees 1 parent GPU and CUDA sees 2 MIG devices.

• Medium cuda_core/tests/system/test_system_device.py:58, cuda_core/cuda/core/system/_device.pyx:367
test_to_cuda_device() now catches any RuntimeError and continues. On the Thor setup from the logs, system.Device.get_all_devices() only yields the physical NVML parent, so this test can make zero assertions and still pass. That masks both the intended MIG-parent exception and any future UUID-mapping regression.

• Medium cuda_bindings/tests/nvml/test_cuda.py:61
  The new Thor handling is an unconditional product-name skip. That disables coverage for every Thor host, not just the specific "CUDA enumerates MIG instances while NVML enumerates the parent GPU" case, and it does nothing for the same topology if it shows up under a different product name.

• Medium cuda_bindings/tests/nvml/conftest.py:103, cuda_bindings/tests/nvml/conftest.py:118, cuda_bindings/tests/nvml/test_pynvml.py:45
mig_handles() still derives the MIG index upper bound from handles[0] and then reuses that bound for every GPU. On mixed systems, if GPU 0 has fewer or zero MIG slots than a later GPU, the fixture becomes incomplete or empty and test_device_get_attributes() skips even though live MIG handles exist.

• Low cuda_core/tests/test_device.py:50, cuda_core/cuda/core/system/_device.pyx:899
  The new pci_info() behavior is effectively untested. contextlib.suppress(RuntimeError) hides all pci_info failures, not just the expected MIG-without-PCI case, so a real regression on a normal device would also pass silently.