benchmarks: cupy branch in get_xr_dataarray never raises NotImplementedError on non-GPU hosts #3707

benchmarks/benchmarks/common.py gates the cupy branch with if not has_cuda_and_cupy: — but has_cuda_and_cupy is a function, so the expression is always false and the guard never fires. On a host without CUDA the branch falls through to import cupy and the benchmark dies with ImportError instead of being skipped via NotImplementedError.

The dask+cupy branch added in #3706 calls the function correctly; this fixes the older branch to match. One-line change plus whatever guard-behavior check is practical without a GPU-less CI lane.

Spotted during the #3706 review; kept out of that PR to leave the diff scoped.