[Code scan] Validate HopDict diagonal data before mutating state

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

Severity: Medium

`HopDict.__setitem__()` copies the matrix into `_hoppings` before checking whether `(0, 0, 0)` contains forbidden diagonal onsite terms. If `_check_diag()` raises and the caller catches the exception, the invalid matrix remains stored in the object.

Code reference:
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/builder/base.py#L542-L548

Suggested fix: validate the candidate matrix before assigning it to `_hoppings`, or roll back the assignment when validation fails.

Start in tbplas/builder/base.py around HopDict.__setitem__() at lines 542-548, then inspect _check_diag() to understand when validation raises. Ensure the candidate matrix is validated before _hoppings changes, and verify that a caught validation error leaves the previous state unchanged.

Written by the indexing model from the issue text.