fix(testing_utils): guard get_device_capability() with torch.cuda.is_available()

[Aftabbs]

What does this PR do?

Fixes #45341.

get_device_properties() in testing_utils.py calls torch.cuda.get_device_capability() whenever IS_CUDA_SYSTEM or IS_ROCM_SYSTEM is True. This raises a RuntimeError on environments where CUDA drivers are installed (so torch.version.cuda is not None) but no physical GPU is attached (e.g., Lightning AI Studio CPU-only instances, CI runners with CUDA drivers but no GPU).

Root cause: IS_CUDA_SYSTEM reflects whether the CUDA toolkit is present, not whether a CUDA-capable device is available at runtime. torch.cuda.get_device_capability() requires an actual device.

Fix: Add and torch.cuda.is_available() to the condition so that get_device_capability() is only called when a CUDA/ROCm device is actually present. When CUDA is installed but no device is available, the function falls through to the generic else branch and returns (torch_device, None, None).

-    if IS_CUDA_SYSTEM or IS_ROCM_SYSTEM:
+    if (IS_CUDA_SYSTEM or IS_ROCM_SYSTEM) and torch.cuda.is_available():

Before submitting

• This PR fixes a bug (non-breaking change that fixes an issue)
• This PR is a new feature (non-breaking change that adds functionality)
• This PR is a breaking change (fix or feature that would cause existing functionality not to work as expected)
• This PR adds tests that prove my fix is effective or that my feature works: N/A — the crash only occurs on systems with CUDA installed but no GPU, which is not a typical CI environment; the fix is a one-line guard matching existing patterns elsewhere in the file (e.g. line 995).

[Rocketknight1]

There was already an open PR doing basically the same thing! Please check for other PRs first before sending your agent to fix issues