`ProjectTo(::AbstractArray)` does not infer

e.g. see

julia> @inferred ProjectTo(rand(3, 3))(Diagonal(rand(3)))
ERROR: return type Diagonal{Float64, Vector{Float64}} does not match inferred return type Union{Base.ReshapedArray{Float64, 2, Diagonal{Float64, Vector{Float64}}, Tuple{Base.MultiplicativeInverses.SignedMultiplicativeInverse{Int64}}}, Diagonal{Float64, Vector{Float64}}}

the reason for the failure is
https://github.com/JuliaDiff/ChainRulesCore.jl/blob/6efb2d258dc225a50d961fb17fa5ba88d7296ce7/src/projection.jl#L174-L183
which decides whether to dy is dx or a reshape(dx, ...) based on the values of the axes.

We could instead do

    dy = if length(axes(dx)) == length(project.axes)
        axes(dx) == project.axes || throw(_projection_mismatch(project.axes, size(dx)))
        dx
    else
        for d in 1:max(M, length(project.axes))
            if size(dx, d) != length(get(project.axes, d, 1))
                throw(_projection_mismatch(project.axes, size(dx)))
            end
        end
        reshape(dx, project.axes)
    end

which does infer, but throws an error for the arrays that need to be reshaped. In practice, this only means that only

        poffv = ProjectTo(OffsetArray(rand(3), 0:2))
        @test axes(poffv([1, 2, 3])) == (0:2,)

test fails in ChainRulesCore
and all the rrule inference tests (but not the one where inplaceablethunk inference fails) in https://github.com/JuliaDiff/ChainRules.jl/pull/459#issuecomment-884757000 are fixed.

do we want this tradeoff or not?

Start with src/projection.jl at the referenced lines and reproduce the @inferred ProjectTo example. Compare the proposed axis-length behavior with the OffsetArray test in ChainRulesCore and the rrule inference tests mentioned in ChainRules#459. Done means deciding the tradeoff and updating the relevant tests and implementation accordingly.