`show` is confused by shared parameters

[mcabbott]

Here it concludes that there are non-trainable arrays, but in fact the same weights appear in two layers, which throws off the counting:

julia> using Flux

julia> let d = Dense(10 => 10)
       Chain(Embedding(10=>10), d, d)
       end
Chain(
  Embedding(10 => 10),                  # 100 parameters
  Dense(10 => 10),                      # 110 parameters
  Dense(10 => 10),                      # 110 parameters
)         # Total: 3 trainable arrays, 210 parameters,
          # plus 2 non-trainable, 110 parameters, summarysize 1.055 KiB.

julia> Flux.destructure(ans)  # length 210, correct
(Float32[1.7007293, 0.66258854, -0.040887665, -1.2084905, -0.53106576  …  0.0, 0.0, 0.0, 0.0, 0.0], Restructure(Chain, ..., 210))