[Code scan] Preserve finite s-orbital cube values at the nucleus

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

Severity: Low

`set_cube()` forces every grid point with `r < 1.0e-7` to zero. That avoids division by zero for angular functions with `l > 0`, but it also zeros `l == 0` s orbitals at the nucleus, where the value is finite. The resulting cube output contains an artificial node at the nucleus for s states.

Code reference:
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/cython/atom.pyx#L193-L209

Suggested fix: handle `l == 0` separately at small `r`, and only force zero for angular components that are genuinely undefined or vanish at the origin.

Start in tbplas/cython/atom.pyx at set_cube(), especially lines 193-209, and inspect how the small-r cutoff interacts with the orbital angular component. Verify the behavior for l == 0 and l > 0; done means finite s-orbital values are preserved at the nucleus while genuinely undefined or vanishing components remain zero.

Written by the indexing model from the issue text.