Type stability regression for `mean(::Array{RGB{N0f8}})` from 0.6 to 0.7 about colors.jl

[kimikage]

Let us sort out the issue.

First of all, Colors.jl does not provide arithmetic operation methods for colors except XYZ and LMS. (cf. Support arithmetic only for linear color spaces)
So, it seems to me that the inferred return type Union{RGB{Float32}, RGB{Float64}} comes from the codes of ColorVectorSpace.jl.

Second, what is the expected return type of mean(::Array{RGB{N0f8}})?
If you mind the type stability because of the processing speed, I wonder whether RGB{Float64} is always best. Moreover, to stop using mean may be a possible measure.
If the problem is not the speed, why not explicitly convert the result into preferred type? For example,

julia> Test.@inferred convert(eltype(A), mean(A))

By the way, the return type of Colors.weighted_color_mean() is the same as the inputs, i.e. RGB{N0f8} in this case, although it may cause a precision issue.

Well, if it is a pure backward compatibility issue, it's time to give up.
Julia v0.7 (v1.0) includes many breaking changes, you know.

[timholy]

@Evizero, sorry that I failed to notice this in the crush of updating packages to Julia 1. Fixed over at JuliaGraphics/ColorVectorSpace.jl#103