Support for Reactant.jl

[mcabbott]

This doesn't work yet, but wants to work like this:

julia> using Flux, Fluxperimental, Reactant, Enzyme
julia> img = rand32(28, 28, 1, 128);
julia> loss(m, x) = sum(abs2, m(x));
julia> mlp = Chain(Flux.flatten, Dense(28^2 => 32, tanh), Dense(32 => 10));

julia> mlp(img)[1:3]  # plain Julia
3-element Vector{Float32}:
  0.22694848
 -0.72605485
  0.57976365

julia> re_mlp = Fluxactor(mlp);  # uses Reactant

julia> re_mlp(img)[1:3]
┌ Info: compiling...
└   summary(xr) = "28×28×1×128 ConcreteRArray{Float32, 4}"
3-element ConcreteRArray{Float32, 1}:
  0.22694828
 -0.72605544
  0.57976323

julia> re_mlp  # after forward but not yet gradient
Fluxactor(
  Chain(
    Flux.flatten,
    Dense(784 => 32, tanh),             # 25_120 parameters
    Dense(32 => 10),                    # 330 parameters
  ),
  # compiled for 28×28×1×128 ConcreteRArray{Float32, 4}
  # norm(∇) ≈ 0.0f0
  # ∇compiled for nothing
)         # Total: 4 trainable arrays, 25_450 parameters,
          # plus 4 non-trainable, 25_450 parameters, summarysize 689 bytes.

julia> Flux.gradient(loss, mlp, img)[1].layers[2].bias[1:3]  # uses Zygote
3-element Vector{Float32}:
   90.490005
 -208.77806
   28.711397

julia> Flux.gradient(loss, re_mlp, Const(img))[1].layers[2].bias[1:3]  # uses Reactant
[ Info: compiling gradient(loss, ::Fluxactor)...