Extra restructure in Enzyme for reinterpreted array

[ChrisRackauckas]

MWE:

using Zygote, SciMLSensitivity, OrdinaryDiffEq

solver = Rodas3() # fails because of view https://github.com/SciML/PreallocationTools.jl/pull/32

p = rand(3)

function dudt(u, p, t)
    u .* p
end

function loss(p, sensealg)
    prob = ODEProblem(dudt, [3.0, 2.0, 1.0], (0.0, 1.0), p)
    sol = solve(prob, solver, dt = 0.01, saveat = 0.1, sensealg = sensealg,
                abstol = 1e-5, reltol = 1e-5)
    sum(abs2, Array(sol))
end

dp1 = Zygote.gradient(p -> loss(p, QuadratureAdjoint(autojacvec = EnzymeVJP())), p)[1]

The commit a669d4f works around SciML/PreallocationTools.jl#32 since Enzyme can only have Enzyme.DuplicatedNoNeed(y, tmp1) both as the same type. This is then required since:

function get_tmp(dc::DiffCache, u::AbstractArray{T}) where {T <: ForwardDiff.Dual}
    nelem = div(sizeof(T), sizeof(eltype(dc.dual_du))) * length(dc.du)
    if nelem > length(dc.dual_du)
        enlargedualcache!(dc, nelem)
    end
    _restructure(dc.du, reinterpret(T, view(dc.dual_du, 1:nelem)))
end

the return for tmp1 is a ReinterpretedArray instead of an Array. This makes the Rosenbrock case have an unnecessary allocation as it should in theory be able to be kept as a ReinterpretedArray, but this fix requires the extra allocation that SciML/PreallocationTools.jl#32 would have otherwise eliminated.

[ChrisRackauckas]

To fix this, I think either escape analysis has to delete the wrapper allocation or Enzyme has to allow non-same types.

@vchuravy @wsmoses could that be possible?