[Code scan] Use absolute onsite energies in the rescale estimate

This issue comes from a Codex global scan of deepmodeling/tbplas at commit 4d3652b3dba8eb930e4869a6d41074f62d703692.

Severity: Medium

`get_rescale()` documents the estimate as `max(sum(abs(ham_dense), axis=1))`, but initializes each row sum with the raw onsite energy. Negative onsite energies therefore reduce the row sum instead of contributing their absolute value, which can underestimate the Hamiltonian scale and can even make the estimated rescale invalid for diagonal-dominated models.

Code reference:
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/cython/sample.pyx#L80-L113

Suggested fix: initialize `ham_sum[io]` with `abs(orb_eng[io])` so the implementation matches the documented absolute row-sum estimate.

Read tbplas/cython/sample.pyx around get_rescale() at lines 80-113, starting with how the row sum is initialized and how the documented absolute row-sum estimate is computed. Update the initialization to use the onsite energy's absolute value, then verify that the implementation matches the estimate described in the issue.

Written by the indexing model from the issue text.