MDRange not working in some test case

[pgrete]

I tried separating the hierarchical flux kernels (testing performance with potentially reduced register usage), see 3aad08b (and related inline PPM 5954a71 and Kokkos math 9d91ebf).

Performance still tbd, but an issue showed up when trying to change the loop pattern.
The default one

    parthenon::par_for(
        DEFAULT_LOOP_PATTERN, "x1 recon", parthenon::DevExecSpace(),
        0, cons_in.GetDim(5) - 1, 0, cons_in.GetDim(4) - 1, kb.s, kb.e, jb.s, jb.e, ib.s - 1, ib.e + 1

works.
The mdrage_tag one does NOT

    parthenon::par_for(
        parthenon::loop_pattern_mdrange_tag, "x1 recon", parthenon::DevExecSpace(),
        0, cons_in.GetDim(5) - 1, 0, cons_in.GetDim(4) - 1, kb.s, kb.e, jb.s, jb.e, ib.s - 1, ib.e + 1

Moreover, the raw Kokkos one also does NOT work

   Kokkos::parallel_for("x1 flux KMDRange",
         Kokkos::MDRangePolicy<Kokkos::Rank<5>>(parthenon::DevExecSpace(),
          {0, 0, kb.s, jb.s, ib.s - 1},
          {cons_in.GetDim(5), cons_in.GetDim(4), kb.e + 1, jb.e + 1, ib.e + 2},
          {1, 1, 1, 1, ib.e + 2 - (ib.s - 1) }),

Not working here refers to that the code compiles and runs, but that the results are incorrect (indicated by triggering the rho <= 0 fail-safe in ConsToPrim.

Tested on JEDI with

Currently Loaded Modules:
  1) Stages/2024     (S)   6) CUDA/12            (g)  11) hwloc/2.9.1          (g)  16) PMIx/4.2.6                21) ncurses/.6.4     (H)  26) Szip/.2.1.1      (H)  31) libffi/.3.4.4 (H)
  2) GCCcore/.12.3.0 (H)   7) NVHPC/24.3-CUDA-12 (g)  12) OpenSSL/1.1               17) NCCL/default-CUDA-12 (g)  22) bzip2/.1.0.8     (H)  27) HDF5/1.14.2           32) Python/3.11.3
  3) zlib/.1.2.13    (H)   8) XZ/.5.4.2          (H)  13) libevent/.2.1.12     (H)  18) UCC/default          (g)  23) cURL/8.0.1            28) libreadline/.8.2 (H)  33) Ninja/1.11.1
  4) binutils/.2.40  (H)   9) libxml2/.2.11.4    (H)  14) UCX-settings/RC-CUDA      19) MPI-settings/CUDA         24) libarchive/3.6.2      29) Tcl/8.6.13
  5) numactl/2.0.16       10) libpciaccess/.0.17 (H)  15) UCX/default          (g)  20) OpenMPI/4.1.6        (g)  25) CMake/3.26.3          30) SQLite/.3.42.0   (H)

Right now, completely unclear what/why/how/... just opening issue to keep track/revisit next year.