Fix: Prevent MPI communicator growth in LCAO matrix output

[AsTonyshment]

What's changed?

This bug was discovered by @ESROAMER when an LCAO RT-TDDFT calculation failed at approximately electronic step 2370 with:

Too many communicators (0/32768 free on this process; ignore_id=0)

LCAO matrix-output code created a temporary BLACS process grid for each serial gather and then immediately called set_serial(). set_serial() overwrote the BLACS context with -1, so the temporary context could no longer be released. Repeated DMR or H/S(R) output during RT-TDDFT therefore consumed MPI communicator contexts until MPI failed.

The affected output paths now construct their temporary Parallel_Orbitals distributions directly with set_serial(). This is sufficient for gatherParallels(), which performs the data transfer independently of a BLACS context. The same redundant pattern was removed from DMR, H/S(R), derivative-H, decomposed-H, DeePKS H(R), and asynchronous-overlap output paths.

There is no change to matrix values, output formats, filenames, output frequency semantics, numerical solver behavior, or INPUT parameters.

Unit Tests and/or Case Tests for my changes

The figure is plotted directly from the PMPI logs. In two-rank, nine-step RT-TDDFT runs, DMR and H/S(R) output previously added 4 and 8 live communicators per electronic step, respectively, or 12 when combined. After the fix, all output cases match the control and finish with created=41, freed=41, live=0 on each rank.

The generated DMR and H/S(R) CSR files are byte-for-byte unchanged.

Governance Notes

• INPUT/docs changes: None.
• Core module impact: The numerical ESolver, HSolver, ElecState, Hamilt, and Psi paths are unchanged. The only operator-side change is in the optional asynchronous-overlap CSR output path.
• Exceptions requested: None.

[mohanchen]

nice fix!