Parallel write with NF90_INDEPENDENT and NF90_UNLIMITED causes NaN in written file

[dongli]

When parallel write variables with NF90_UNLIMITED dimension, the default access mode NF90_INDEPENDENT causes NaN in written file, but variables without NF90_UNLIMITED dimension are normal. When switched to NF90_COLLECTIVE, both cases are normal, but it is very slow when the data and the process number is very large (e.g., 7200 cores). The NetCDF version is netcdf-c 4.9.2 and netcdf-fortran 4.6.1. I tested in both Linux and macOS with the same behavior. The compiler is gfortran 12.3.0 and openmpi 5.0.0.