Modify the nansum and nanmean function

[Internal-Tide]

The nansum and nanmean functions are no need to copy, it will slow down the code.
This is added new function

function nanmean(x::AbstractArray{T}) where {T<:Number}
    # n = length(x)
    sum = zero(T)
    count = 0
    for i in eachindex(x)
        if !isnan(x[i])
            sum += x[i]
            count += 1
        end
    end
    return sum / count
end

function nanmean(x::AbstractArray{T}, dims) where {T<:Number}
    return mapslices(nanmean, x; dims=dims)
end


function nansum(arr::AbstractArray{T}) where {T<:Number}
    return (sum(x -> !isnan(x) * x, arr))
end

function nansum(arr::AbstractArray{T}, dims) where {T<:Number}
    return (sum(x -> !isnan(x) * x, arr, dims=dims))
end

The speed test is:

using PhysOcean
using BenchmarkTools
using Random

seed = Random.seed!(1234)
a = rand(seed, [NaN, 1.0, 2.5, 0.11, 20.0], (1000, 1000))
@btime nansum($a, 1)
@btime PhysOcean.nansum($a, 1)
@btime nanmean($a, 1)
@btime PhysOcean.nanmean($a, 1)
result: 
110.660 μs (1 allocation: 7.94 KiB)
1.648 ms (11 allocations: 9.29 MiB)
1.740 ms (9501 allocations: 243.50 KiB)
1.889 ms (16 allocations: 9.42 MiB)

The new nansum function is much faster!
However, when the array size is small, the old function is faster than new function, but when the array size is small, I don't think everyone is sensitive to 100 or 200 ns.

[Alexander-Barth]

Thanks a lot!