Change row lengths unit tests to use cpu and single gpu

[edknv]

Follow-up to #113.

This updates test_row_lengths_to_offsets_device to only test cpu <-> cuda:0 and removes the cpu <-> cuda:1 case because the test will fail if we only have a single GPU. We simply remove the second gpu case because testing the cpu <-> cuda:0 case should be sufficient for testing if the tensor concatenation works across devices.