diagm does not correctly preserve eltype for StaticArrays

I think this is a bug, although it may be that for some reason this behavior is expected:

With a regular array, diagm preserves the element type:

julia> using LinearAlgebra, Unitful

julia> v = [ 1.0, 1.0 ]u"nm"
2-element Vector{Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}}:
 1.0 nm
 1.0 nm

julia> eltype(v)
Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}

julia> m = diagm(v)
2×2 Matrix{Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}}:
 1.0 nm  0.0 nm
 0.0 nm  1.0 nm

julia> eltype(m)
Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}

But with static arrays it does not:

julia> using Unitful, LinearAlgebra

julia> using StaticArrays

julia> v = SVector(1.0, 1.0)u"nm"
2-element SVector{2, Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}} with indices SOneTo(2):
 1.0 nm
 1.0 nm

julia> eltype(v)
Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}

julia> m = diagm(v)
2×2 SMatrix{2, 2, Quantity{Float64}, 4} with indices SOneTo(2)×SOneTo(2):
 1.0 nm     0.0
    0.0  1.0 nm

julia> eltype(m)
Quantity{Float64}

Probably irrelevant, but this is how I had implemented that as a workaround:

julia> function my_diagm(v::StaticVector{N}) where {N}
           T = promote_type(typeof.(v)...)
           cart_idxs = CartesianIndices((1:N,1:N))
           D = SMatrix{N,N,T,N*N}( 
               ntuple(N*N) do i
                   c = cart_idxs[i]
                   if c[1] == c[2] 
                       return v[c[1]] 
                   else
                       return zero(T)
                   end
               end
           )
           return D
       end
my_diagm (generic function with 1 method)

julia> v = SVector(1.0, 1.0)u"nm"
2-element SVector{2, Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}} with indices SOneTo(2):
 1.0 nm
 1.0 nm

julia> my_diagm(v)
2×2 SMatrix{2, 2, Quantity{Float64, 𝐋, Unitful.FreeUnits{(nm,), 𝐋, nothing}}, 4} with indices SOneTo(2)×SOneTo(2):
 1.0 nm  0.0 nm
 0.0 nm  1.0 nm

This function will fail if the type promoted has not zero defined, but that seems reasonable to me, as it would be ambiguous what to put off the diagonals.

Reproduce the regular-array and StaticArrays examples in the issue, including the Unitful-valued SVector, then locate the implementation of diagm used for static vectors. Confirm how the off-diagonal zero values determine the result type; done means the static result preserves the input element type and the regression is covered by a test.