`nomissing` function about divand.jl

[Alexander-Barth]

I corrected the docs of NCDatasets, thanks! But I prefer to keep the function nomissing for now because it handles better the case where the data array is Float32 and the user uses a Float64 NaN.
The end-goal, is to be able to use missing everywhere (and in PyPlot in particular, JuliaPy/PyCall.jl#616)

julia> nomissing([1.0f0,2.0f0,missing],NaN)
3-element Array{Float32,1}:
   1.0
   2.0
 NaN  

julia> coalesce.([1.0f0,2.0f0,missing],NaN)
3-element Array{AbstractFloat,1}:
   1.0f0
   2.0f0
 NaN