use MPI file function to output KS_eigenvector

[userchba]

Summary

Replace the rank-0 gather-and-write implementation of the LibRPA v1 KS eigenvector output (KS_eigenvector_0.dat, active when out_librpa_reader_version == 1 under __MPI) with a collective MPI-IO writer.

Motivation

The old implementation had several scalability and robustness problems:

• It issued one MPI_Allreduce (over MPI_COMM_WORLD) per (k-point, spin, band), i.e. nks × nspin × nbands collectives of size nlocal each.
• It materialized the full nbands × nlocal wavefunction matrix on every rank, and rank 0 held the entire file payload in memory (nks × nspin × nbands × nlocal × 16 B) before writing it serially — prohibitive for production RPA/GW system sizes.
• It wrote directly to the final file name, so a mid-write failure could leave a truncated/corrupt KS_eigenvector_0.dat.
• MPI return codes were unchecked; a rank-0 file-open failure threw while other ranks proceeded into the next collective (hang/abort hazard), and the Allreduce was hardcoded to MPI_COMM_WORLD instead of the wavefunction communicator.

What this PR does

RPA_LRI::out_eigen_vector (v1 branch, MPI build) now uses a collective MPI-IO writer:

• Each rank's block-cyclic ownership is described by a run-length-encoded MPI_Type_create_hindexed file datatype; payload is written with MPI_File_write_at_all in chunks bounded to 4M complex elements (64 MB pack buffer per rank, independent of system size).
• Header + k-point directory are written by rank 0 within the same collective calls; no payload Allreduces remain.
• Atomic publication: the file is written to KS_eigenvector_0.dat.tmp, synced, closed, then renamed by rank 0; the temp file is removed on any error.
• Collective error propagation (Allreduce/Bcast-based checks) so all ranks fail together with a clear message; MPI error codes are converted to readable exceptions.
• Collective validation before writing: psi vs Parallel_Orbitals dimensions, spin/SOC consistency (even nlocal, even/odd spinor component split), and an exact global ownership invariant (summed per-rank element counts must equal nbands × spatial_basis).
• All offset arithmetic is overflow-guarded (MPI_Aint / MPI_Offset / int32 range checks).

Compatibility

• The binary v1 format, the file name, and the payload layout ([component][band][orbital], SOC split into even/odd spinor components) are unchanged — downstream LibRPA readers are unaffected.
• The serial (non-MPI) path and the legacy text output (out_librpa_reader_version != 1) are untouched.

Verification

• A standalone MPI test harness including the new writer verbatim from RPA_LRI.hpp, compared against a re-implementation of the old writer: outputs are byte-for-byte identical over ~150 runs spanning 1–8 MPI ranks, nspin = 1/2/4 (incl. SOC), ScaLAPACK block sizes 1–33, uneven grids with ranks owning zero bands, and payloads crossing the 4M-element chunk boundary (multi-chunk collective writes).
• Every produced file was additionally validated against an independent analytic ground truth (header fields, directory offsets, and every payload element).
  ``