[bsr_add] fix the cuda bsr add path and add the cpu one

[zitongzhan]

• Inspect failing CI runs for this branch and confirm root cause
• Revert the CPU execution expansion in tests/autograd/test_bal_jacobian.py by restoring the CUDA-only guards
• Revert the device-parametrize expansion in tests/sparse/test_bsr.py back to CUDA-only
• Verify the working tree only contains the intended reverts

[zitongzhan]

@copilot This PR makes it possible for CPU to perform the BSR add operation. Therefore, many test verifying the correctness of Jacobian that was skipped on non-CUDA device can be relaxed and added back. Identify them.

[Copilot]

@copilot This PR makes it possible for CPU to perform the BSR add operation. Therefore, many test verifying the correctness of Jacobian that was skipped on non-CUDA device can be relaxed and added bac...

Identified and addressed in commit 0d21a05:

tests/autograd/test_bal_jacobian.py — removed two if device.type != "cuda": return early-exit guards:

• test_bal_jacobian_structure_no_empty_columns: the end-to-end LM optimization + per-pixel error check now runs on CPU too
• test_bal_jacobian_cat_fixed_first_camera_gauge_free: same for the fixed-first-camera variant

tests/sparse/test_bsr.py — extended the test_universal device parametrize from ['cuda'] to ['cpu', 'cuda'], with a skip for torch.matmul on CPU (BSR matmul still requires CUDA/warp kernels) and a skip when CUDA is unavailable. This lets torch.add, torch.diag, and diag_clamp_oop all run on CPU.