[Code scan] Clear degenerate AC products in the Cython Lindhard backend

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

Severity: Medium

The Cython `prod_ac()` path assigns `prod_df[ik, mm, nn]` only when the band-energy difference exceeds the threshold. Degenerate entries are left at whatever value the caller supplied. The Fortran implementation explicitly initializes `prod_df = 0.0` before filling, so the two backends have different safety contracts.

Code references:
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/cython/lindhard.pyx#L359-L375
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/fortran/lindhard.f90#L188-L193

Suggested fix: explicitly set the degenerate `prod_df` entries to zero, or initialize the whole output array inside the Cython function before filling it.

Start with the Cython prod_ac() path in tbplas/cython/lindhard.pyx around lines 359-375, then compare its initialization behavior with tbplas/fortran/lindhard.f90 around lines 188-193. Ensure degenerate prod_df entries are zero rather than caller-provided values, and verify that the Cython and Fortran backends have the same safety contract.

Written by the indexing model from the issue text.