[FR] Mixed eltype dot products

These don't seem to currently work (CUDA 3.3):

CuVector{Float32}(undef,10) ⋅ CuVector{Complex{Float32}}(undef,10)
CuVector{Float32}(undef,10) ⋅ CuVector{Float64}(undef,10)
# other mixed eltypes, etc...

as they fall back to a generic which triggers scalar indexing. It would be nice to have these implemented, even with a simple sum(conj.(x) .* y) or something, which at least works on GPU.

Maybe worth mentioning, but I didn't really care about these until upgrading Zygote 0.6.11 -> 0.6.12, since a recent change (bisected down to https://github.com/FluxML/Zygote.jl/pull/973) seems to make it so Zygote emits such dot products where previously it wasn't. Here's an example which triggers scalar indexing after that commit but not before:

using CUDA, Zygote, LinearAlgebra
CUDA.allowscalar(false)

x = cu(rand(10))
y = complex(cu(rand(10)))

Zygote.gradient(1) do A
    norm(Diagonal(A .* x) * y)
end

Anyway, this isn't really relevant for CUDA.jl, but figured might provide some context.

Start by tracing the dot-product dispatch for mixed-eltype CuVector operands, using the Float32/Complex{Float32} and Float32/Float64 examples in the issue. Run the provided Zygote gradient example with CUDA.allowscalar(false); done means these mixed dot products execute on the GPU without triggering scalar indexing.