Fix OpenMP reproducibility

[DusanJovic-NOAA]

Fix the reproducibility of model outputs when a threaded build runs on a different number of OpenMP threads.

I used thread sanitizer to find potential data races, which frequently cause model non-reproducibility. Not every data race results in non-reproducible output, for example, a data race where multiple threads write the same initial value to a shared memory location. I think those cases should be fixed as well to minimize the number of warnings the thread sanitizer reports.

I tested these code changes using Intel (2025.3.0) and GNU (11.5.0). To get reproducible results with optimized code (-O3) for the Intel compiler I had to use the -fp-model precise compiler flag, in addition to the flags currently specified in the CMake files. For the GNU compiler I had to use -fno-tree-loop-vectorize. I didn't add those flags to the CMake files because I haven't checked the performance impact.

With the code changes included in this PR and the compiler flags mentioned above, I am getting reproducible outputs (history, diag and restart files, checked by running nccmp) when I use 'mesoscale_reference' physics. For 'convection_permitting' physics two additional changes are needed in physics_mmm:

diff --git a/bl_mynn.F90 b/bl_mynn.F90
index 783b996..c083888 100644
--- a/bl_mynn.F90
+++ b/bl_mynn.F90
@@ -312,7 +312,7 @@
  integer,intent(out):: &
     errflg        ! output error flag (-).
 
- real(kind=kind_phys),intent(out),dimension(:):: &
+ real(kind=kind_phys),intent(out),dimension(its:):: &
     maxwidth,   &!
     maxmf,      &!
     ztop_plume

and

diff --git a/cu_ntiedtke.F90 b/cu_ntiedtke.F90
index e1d266d..75367e5 100644
--- a/cu_ntiedtke.F90
+++ b/cu_ntiedtke.F90
@@ -13,6 +13,7 @@
  real(kind=kind_phys):: g
  real(kind=kind_phys):: rd
  real(kind=kind_phys):: rv
+ !$omp threadprivate(alf, als, alv, cpd, g, rd, rv)
 
  real(kind=kind_phys),parameter:: t13   = 1.0/3.0
  real(kind=kind_phys),parameter:: tmelt = 273.16
@@ -46,6 +47,7 @@
  real(kind=kind_phys):: ralfdcp
  real(kind=kind_phys):: vtmpc1
  real(kind=kind_phys):: zrg
+ !$omp threadprivate(rcpd, c2es, c5les, c5ies, r5alvcp, r5alscp, ralvdcp, ralsdcp, ralfdcp, vtmpc1, zrg)
 
  logical,parameter:: nonequil = .true.
  logical,parameter:: lmfpen   = .true.