fix: Improve CUDA version detection and error handling

[ved1beta]

fix: Improve CUDA/HIP version detection and error handling

This commit fixes issue #1513 by improving the version detection logic
in cuda_specs.py to handle both CUDA and ROCm/HIP systems more robustly.

Key changes:

• Add proper error handling for None values in version detection
• Make version string and tuple functions return Optional types
• Add robust version string parsing with try-except blocks
• Add validation checks for compute capabilities
• Improve error handling in get_cuda_specs function
• Add proper type hints and documentation

The changes ensure that:

1. NoneType errors are properly handled when version info is missing
2. ROCm/HIP systems are properly supported
3. Invalid version strings don't cause crashes
4. All error cases return None instead of raising exceptions

Test results show successful version detection for both CUDA and ROCm/HIP
systems, resolving the original issue where torch.version.cuda.split()
would fail on ROCm systems.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

LGTM, thanks! We can merge after fixing the failing linter check.