`promote_type` on GPU arrays of Float32 and ComplexF32 promotes to UnionAll

Hello,

I stumbled over this difference between CPU and GPU arrays:

```julia
julia> using JLArrays

julia> using CUDA

julia> a = rand(Float32, 1); b = rand(ComplexF32, 1);

julia> promote_type(typeof(a), typeof(b))
Vector{ComplexF32} (alias for Array{Complex{Float32}, 1})

julia> promote_type(typeof(JLArray(a)), typeof(JLArray(b)))
JLArray{T, 1} where T

julia> promote_type(typeof(CuArray(a)), typeof(CuArray(b)))
CuArray{T, 1, CUDA.DeviceMemory} where T
```

I've noticed this behaviour originally on CuArrays, but noticed it is also present for JLArrays, so I hope this is the correct repository. I have not been able to test it out on other GPU arrays (yet)

Start by reproducing the promote_type results shown for CPU arrays, JLArray, and CuArray, then inspect the GPUArrays.jl array type and promotion entry points. A fix should make Float32 and ComplexF32 GPU array promotion produce a concrete promoted array type rather than a UnionAll, with matching regression coverage.

Written by the indexing model from the issue text.