Issues with degenerate input for pooling layers

[staticfloat]

Our pooling layers don't deal well with degenerate inputs (e.g. ones()). MWE:

using Flux
x = ones(10, 10, 1, 1)
xp = param(x)
y_hat = MaxPool((3,3), stride=(2,2))(xp)
Flux.back!(sum(y_hat))

Which results in:

julia> xp.grad
10×10×1×1 Array{Float64,4}:
[:, :, 1, 1] =
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 2.0  2.0  4.0  2.0  4.0  2.0  4.0  2.0  2.0  0.0
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 2.0  2.0  4.0  2.0  4.0  2.0  4.0  2.0  2.0  0.0
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 2.0  2.0  4.0  2.0  4.0  2.0  4.0  2.0  2.0  0.0
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 1.0  1.0  2.0  1.0  2.0  1.0  2.0  1.0  1.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

Compare with adding a bit of noise to eliminate the degeneracy:

xp = param(x .+ 0.01.*randn(size(x)...))
y_hat = MaxPool((3,3), stride=(2,2))(xp)
Flux.back!(sum(y_hat))

Which results in the proper output of prod(size(y_hat)) == sum(xp.grad) :

julia> xp.grad
10×10×1×1 Array{Float64,4}:
[:, :, 1, 1] =
 0.0  0.0  2.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  1.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  1.0  0.0  0.0  0.0  0.0  2.0  0.0
 0.0  0.0  0.0  0.0  1.0  0.0  0.0  0.0  0.0  0.0
 0.0  2.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  1.0  0.0  0.0  0.0  1.0  0.0
 0.0  0.0  0.0  1.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0
 0.0  0.0  2.0  0.0  0.0  1.0  0.0  0.0  1.0  0.0
 0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0  0.0

[staticfloat]

Fixed by #94