Summary
Convert 8 tests verified to have no CUDA-specific code in body from @onlyCUDA to @onlyAccelerator, making them runnable on XPU, HPU, and PrivateUse1 backends.

6 clean swaps: test_norm_fused_type_promotion, test_powsum_dtype_kwarg, test_foreach_powsum_dtype_kwarg, test_triangular_solve_large, test_lu_solve_large_matrices,...