Migrate NvmlGpuInfo from pynvml to cuda.core.system after 1.0 release

[mmccarty]

Context

PR #137 introduced NvmlGpuInfo in hardware.py which wraps raw pynvml calls behind a GpuInfoProvider Protocol. Reviewers (@ncclementi, @mdboom) noted we should migrate from pynvml to cuda.core.system for consistency with the rest of the CUDA Python ecosystem.

As of recently, cuda.core.system has all the APIs needed (device count, compute capability, memory info, driver version). However, this requires waiting for the cuda-core 1.0 release before merging.

Related

• Migrate pynvml to cuda.core.system #146 — draft PR by @mdboom migrating pynvml to cuda.core.system (blocked on cuda-core 1.0 release)
• cuda_toolkit_check fails on fresh install as cuda-bindings is not declared in dependencies #145 — cuda-bindings dependency issue
• Refactor hardware access with dependency injection #137 — DI refactor that introduced the GpuInfoProvider abstraction

Work needed

Once cuda-core 1.0 is released:

1. Update NvmlGpuInfo in rapids_cli/hardware.py to use cuda.core.system instead of pynvml
2. Update HardwareInfoError wrapping to catch cuda.core exceptions instead of pynvml.NVMLError
3. Remove nvidia-ml-py (pynvml) from dependencies.yaml if fully replaced
4. Update tests in test_hardware.py to mock cuda.core instead of pynvml