Gemini3d

fix segfault on large e.g. 600x600x600 simulations by fixing MatGemini to write proper scalar datasets, and h5fortran to verify scalar is read into scalar as array into scalar segfaults.

Also switched allocatable actual args to have intent(inout) dummy args, as forced reallocation results otherwise, which is unintended but Fortran standard 2003 behavior.

* Added our own FindMPI as the vendor FindMPI is frequently a source of trouble as like vendor FindHDF5, it covers a large range of rare/old systems, and fails to work on our more up to date systems
* ci: added JUnit dash demo
* Find*.cmake: works better with CentOS packages
* numerous magcalc improvements, and add magcalc.run frontend
* remove Apple Silicon M1 workaround in anticipation of HWLOC 2.6, which we contributed to handling this and future Apple Silicon properly.

As before, Gemini3D works with GCC 8, 9, 10, 11. Gemini3D also now works with GCC 7.5.0.
GCC older than 7.5.0 is unlikely to work. For example GCC 7.4.0 is known to be broken.

Updated h5fortran,nc4fortran interfaces to auto-close files on leaving scope--still recommended to use %close() to be clear.

Fixed/updates Matlab self-tests, although Python is recommended as more complete as faster as before.

improve quality of .run frontends with functionalization and reuse

* gemini3d.run
* magcalc.run

* patched Zlib 2.0.3 (will be included in next Zlib-ng release) to work with GCC mpicc
* make h5fortran internals TKR interface procedure instead of `select type` for robustness and runtime improvements