Inconsistent interpolation outcome on Float32 inputs

[cstjean]

When x::Vector{Int} and y::Vector{Float32}, I expect the outcome to be a Float32, but it's inconsistent:

julia> interp = Interpolations.interpolate(([1,2,3],), [1.0f0, 2.0f0, 1.0f0], Interpolations.Gridded(Interpolations.Linear()));

julia> interp(2)
2.0

julia> interp([2])
1-element Array{Float64,1}:
 2.0

julia> ex = Interpolations.extrapolate(interp, NaN32);

julia> ex(2)
2.0

julia> ex([2])
1-element Array{Float32,1}:
 2.0

If my inputs are Float32, then all is fine, but for my use case, x is the number of seconds since some reference time, and I am a bit worried about the loss of precision that comes from early Float32 conversion.

[stevengj]

I am a bit worried about the loss of precision that comes from early Float32 conversion.

Note that there is no loss of precision in converting an Int to Float32 if your integers are ≤ maxintfloat(Float32) == 16777216.

That being said, I agree that it should use float(promote_type(Int, Float32)) == Float32. @timholy noted in #359 that the promotion code in this package needs some cleanup.