NetCDF output with BFloat16

[milankl]

fixes #406, T21 simulation with BFloat16!

out.mp4

There remains an issue with the restart files

ERROR: MethodError: no method matching round!(::LowerTriangularMatrix{Complex{BFloat16}}, ::Int64)

Closest candidates are:
  round!(::AbstractArray{T}, ::Integer) where T<:Union{Float16, Float32, Float64, ComplexF64, ComplexF16, ComplexF32}
   @ BitInformation ~/.julia/packages/BitInformation/CkrSa/src/round_nearest.jl:90

which however should probably be fixed in BitInformation.jl (at the moment it just doesn't create a restart file, but netcdf is fine (see video)

@madnanabid

[milankl]

@croci some vorticity advection with a pseudo-spectral solver in BFloat16! At the moment going to higher resolution fails though, probably something wrong with the spectral transform!? Could be interesting to test out BFloat16sr in comparison!