scatter cleanup and fix min/max

[CarloLucibello]

Fix #317. The main problem was in the use of FiniteDifferences' central difference method on singularities of non-smooth functions.

[yuehhua]

I am quite curious about the semantic difference between DoesNotExist and NoTangent.

[CarloLucibello]

I am quite curious about the semantic difference between DoesNotExist and NoTangent.

no difference, just a recent renaming in ChainRulesCore.jl

[yuehhua]

Maybe we could come out with a new name for _check_dims? Maybe _check_scatter_dims or something else?

[yuehhua]

For scatter, I have been used to write something like this.

scatter(op,
        src::AbstractArray{Tsrc,Nsrc},
        idx::AbstractArray{Tidx,Nidx};
        init = scatter_empty(op, Tsrc))

I think it can avoid branching on isnothing.

[yuehhua]

The dimension checking is slightly different for scatter and gather. scatter calls _check_dims/scatter_dims with (dst, src, idx), while, reversely, gather calls _check_dims/scatter_dims with (src, dst, idx). For abstraction, I give _check_dims/scatter_dims a signature of (X, Y, idx), so they are both available for scatter and gather.

[yuehhua]

I am good with this PR.

[CarloLucibello]

buildkite fail is unrelated, merging