Don't use the default value for device fixture as this will break it. (#7675)

Example for XPU backend:
```bash
=========================================================================== short test summary info ============================================================================
FAILED python/test/unit/language/test_core.py::test_split_subview[128-128-64-64] - AssertionError: Torch not compiled with CUDA enabled
FAILED python/test/unit/language/test_core.py::test_split_subview[128-128-64-32] - AssertionError: Torch not compiled with CUDA enabled
FAILED python/test/unit/language/test_core.py::test_split_subview[128-64-64-32] - AssertionError: Torch not compiled with CUDA enabled
FAILED python/test/unit/language/test_core.py::test_split_subview[256-128-64-64] - AssertionError: Torch not compiled with CUDA enabled
============================================================================== 4 failed in 4.99s ===================================================
```

Author: anmyachev

python/test/unit/language/test_core.py

@@ -6515,7 +6515,7 @@ def test_local_load_store(M, N, K, dist_layout, shared_layout, device, tmp_path:
 
 @pytest.mark.parametrize("M, N, M_tile_size, N_tile_size",
                          [[128, 128, 64, 64], [128, 128, 64, 32], [128, 64, 64, 32], [256, 128, 64, 64]])
-def test_split_subview(M, N, M_tile_size, N_tile_size, device='cuda'):
+def test_split_subview(M, N, M_tile_size, N_tile_size, device):
     num_rows_per_warp = THREADS_PER_WARP // 4
     num_repeats_M = triton.cdiv(M, M_tile_size)
     num_repeats_N = triton.cdiv(N, N_tile_size)