Refactor hardware access with dependency injection

[mmccarty]

Summary

• Add rapids_cli/hardware.py with provider abstractions: DeviceInfo dataclass, GpuInfoProvider/SystemInfoProvider protocols, NvmlGpuInfo/DefaultSystemInfo real implementations (lazy-loading, cached), and FakeGpuInfo/FakeSystemInfo/FailingGpuInfo/FailingSystemInfo test fakes
• Refactor gpu.py, cuda_driver.py, memory.py, nvlink.py, and debug.py to accept optional provider parameters instead of calling pynvml/psutil/cuda.pathfinder directly
• Update doctor.py to create a shared NvmlGpuInfo instance and pass it to all checks
• Rewrite check/debug tests to use plain dataclass fakes instead of mock.patch — eliminates ~51 hardware patches and ~11 MagicMock objects from those tests
• Add test_hardware.py with 19 tests for all provider implementations
• Fix nvlink bug: nvmlDeviceGetNvLinkState(handle, 0) always passed 0 instead of actual nvlink_id

Note: This branch includes commits from #135 (tests) and #136 (docs). The DI-specific changes are in commit 28e3957.

Test plan

• pytest — all 72 tests pass
• Coverage at 97.72%, above 95% threshold
• pre-commit run --all-files passes

🤖 Generated with Claude Code