[Code scan] Pardiso HDF5 export fails for CUDA tensors

This issue is part of a Codex global repository scan.

Problem:
`TBSystem._save_h5()` writes tensor blocks with `block.detach().numpy()`. Pardiso export gets HR/SR from the calculator on the model device, so CUDA-backed models can pass CUDA tensors into this writer.

Code reference:
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/postprocess/unified/system.py#L902-L907

Impact:
`to_pardiso()` and `to_pardiso_json()` can fail during HDF5 export when the model/calculator runs on CUDA.

Suggested fix:
Use `block.detach().cpu().numpy()` for tensor blocks before creating HDF5 datasets.

Open dptb/postprocess/unified/system.py around TBSystem._save_h5() at lines 902-907 and trace how to_pardiso() and to_pardiso_json() reach the HDF5 export. Verify the failure with CUDA-backed tensor blocks, then confirm both export paths complete and produce readable HDF5 output without the CUDA conversion error.

Written by the indexing model from the issue text.