sensitivity of sum with dims has trailing singlton dimension

[oxinabox]

I don't know if this was introduced in #123 or merely uncovered and tested for.
But the tests in #123 are wrong AFAICT.

Consider:

julia> z = [1.0, 2.0, 3.0];

julia> size(z)
(3,)

julia> z̄, = ∇(x->sum(sum(x'; dims=1)))(z)
([1.0; 1.0; 1.0],)

julia> size(z̄)
(3, 1)

it is an intrinstict truth of reverse mode AD that the structure of z and of z̄ must be the same.

[oxinabox]

I can't work out the true cause, but this seems to be fixed in ChainRules and will be closed by #189