@non_differentiable _denom

[mcabbott]

Fixes this:

julia> Zygote.hessian_reverse(x -> sum(sin, mean(x.^2; dims=1)), [1 2; 3 4.0])  # Int
4×4 Matrix{Float64}:
 1.24259  2.87677  0.0      0.0
 2.87677  8.91398  0.0      0.0
 0.0      0.0      1.33701  4.35217
 0.0      0.0      4.35217  7.86527

julia> Zygote.hessian_reverse(x -> sum(sin, mean(x.^2; dims=[1])), [1 2; 3 4.0])  # Vector
4×4 Matrix{Float64}:
 1.24259  2.87677  0.0      0.0
 2.87677  8.91398  0.0      0.0
 0.0      0.0      1.33701  4.35217
 0.0      0.0      4.35217  7.86527

julia> Zygote.hessian_reverse(x -> sum(sin, mean(x.^2; dims=(1,))), [1 2; 3 4.0])  # Tuple
ERROR: Mutating arrays is not supported -- called push!(Vector{Int64}, ...)
...
Stacktrace:
...
  [5] unique
    @ ./set.jl:176 [inlined]
  [6] (::typeof(∂(unique)))(Δ::Nothing)
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface2.jl:0
  [7] _denom
    @ ~/.julia/packages/ChainRules/hVHC4/src/rulesets/Statistics/statistics.jl:7 [inlined]
  [8] (::typeof(∂(_denom)))(Δ::Float64)
    @ Zygote ~/.julia/packages/Zygote/SmJK6/src/compiler/interface2.jl:0
  [9] #rrule#1801
    @ ~/.julia/packages/ChainRules/hVHC4/src/rulesets/Statistics/statistics.jl:13 [inlined]
 [10] (::typeof(∂(#rrule#1801)))(Δ::Tuple{Matrix{Float64}, NamedTuple{(:n, :sum_pullback), Tuple{Float64, Nothing}}})

Edit: Also adds some Zero paths which seem necc. for FluxML/Zygote.jl#1328.

And a trivial foreach case, for FluxML/Zygote.jl#1297

[oxinabox]

@mcabbott per colprac

PRs must have 1 approval before they are merged.

You can always ping me on Julia slack if you ware waiting for approval on something.
(I get too many github notifications right now to always keep up with them)

[oxinabox]

This seems fine though
Ideally we wouldn't need those zero paths, since Zygote should be dropping them anyway, but its no big deal to have them