Replace usage of CudaStreamDevice with GpuStreamDevice in tensor benchmarks GPU

This MR addresses an issue in the tensor gpu benchmarks where the code was incorrectly utilizing CudaStreamDevice instead of GpuStreamDevice.