Add RIKYU support for MHDTurbulence

[william-dawson]

What

• RIKYU rows in list.csv: 1, 4, and 8 GPUs (1×1, 1×4, 2×4), 10-min budget, as-shipped config.
• RIKYU build/run cases: nvhpc-hpcx-cuda13/26.5; build.sh produces one binary per rank count (Simulation.x.t{1,4,8}) because ntiles is a compile-time parameter; run.sh selects by nodes × numproc_node and launches through a per-rank GPU wrapper.
• Two portability patches (grep-guarded in build.sh, no-ops once upstream):
  • random_seed_put_size.patch — random_seed(PUT=) needs the compiler's seed size (8 for gfortran); physically inert (rrv=0).
  • acc_init_before_mpi.patch — initialize the CUDA context before MPI_Init.
• Launch wrapper (RIKYU run.sh): CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK per rank.

Why the patch + wrapper are required together

Without them, multi-node GPU runs hang silently in the first BoundaryCondition: UCX fails the device-buffer MPI_ISEND with cannot find remote protocol for: inter-node tag_send(multi) from cuda/GPU, and the code waits forever in MPI_WAITALL (the app sets MPI_ERRORS_RETURN and never checks the ISEND return). The context must exist before MPI_Init and each rank must pin exactly one GPU at launch; either part alone still fails. Excluding CUDA transports is not an alternative — hpcx-2.50 has no host-staging path for device buffers (MPI_INIT aborts, "no copy across memory types transport").

Validation (2026-09-03, Rikyu)

• 8 GPUs / 2 nodes: completed clean, 13,764 steps, dt bit-identical (14 digits) to the 1-GPU run; wall 57.1 s vs 115.6 s at 1 GPU.
• Patches originate from william-dawson/MHDTurbulence branch gfortran-port (commits d9a19d9, 4db3adb); candidate upstream PRs to cfcanaoj/MHDTurbulence.

Draft for review of the build/run structure and the patch-application pattern.