Error handling in CUDA kernel launch