Raise an error on import for unsupported GPUs.

[bdice]

Description

RAPIDS 24.02 dropped support for Pascal GPUs. When using an unsupported GPU, the behavior of cudf is undefined and sometimes produces results that appear valid (and empty) but conceal CUDA kernel launch errors. This PR changes the behavior to error on import if unsupported GPUs are detected.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[bdice]

I tested this locally. If you run on a Pascal (unsupported) GPU, you get:

cudf.errors.UnsupportedCUDAError: A GPU with NVIDIA Volta™ (Compute Capability 7.0) or newer architecture is required.
Detected GPU 0: Tesla P100-SXM2-16GB

Detected Compute Capability: 6.0

If you run on a CPU-only machine, or set CUDA_VISIBLE_DEVICES="" on a machine with GPUs (supported or not), you get UserWarning: No NVIDIA GPU detected.

If users need to run on a Pascal machine as-if it is a CPU-only machine, setting CUDA_VISIBLE_DEVICES="" is a sufficient workaround.

[bdice]

@shwina Please see comment above and merge if you're good with that behavior.

[shwina]

I'll note here that RAPIDS_NO_INITIALIZE=1 is another way to circumvent this check if needed.

[shwina]

/merge